Mathematical Modeling and Models For Optimal Decision-Making in Health Care

Computational and Mathematical Methods in Medicine
Mathematical Modeling and Models

for Optimal Decision-Making in
Health Care
Lead Guest Editor: Giedrius Vanagas
Guest Editors: Tomas Krilavičius and Ka L. Man
Mathematical Modeling and Models for Optimal
Decision-Making in Health Care

Lead Guest Editor: Giedrius Vanagas

Guest Editors: Tomas Krilavicius and Ka L. Man
Copyright © 2019 Hindawi. All rights reserved.
This is a special issue published in “Computational and Mathematical Methods in Medicine.” All articles are open access articles dis-
tributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium,
provided the original work is properly cited.
Editorial Board
Raul Alcaraz, Spain Luca Faes, Italy Kemal Polat, Turkey
Emil Alexov, USA Maria E. Fantacci, Italy Alberto Policriti, Italy
Konstantin G. Arbeev, USA Giancarlo Ferrigno, Italy Giuseppe Pontrelli, Italy
Georgios Archontis, Cyprus Marc Thilo Figge, Germany Jesús Poza, Spain
Enrique Berjano, Spain Alfonso T. García-Sosa, Estonia Christopher Pretty, New Zealand
Junguo Bian, USA Humberto González-Díaz, Spain Mihai V. Putz, Romania
Elia Biganzoli, Italy Igor I. Goryanin, Japan Jose Joaquin Rieta, Spain
Konstantin Blyuss, UK Marko Gosak, Slovenia Jan Rychtar, USA
Hans A. Braun, Germany Damien Hall, Australia Vinod Scaria, India
Zoran Bursac, USA Roberto Hornero, Spain Simon A. Sherman, USA
Thierry Busso, France Tingjun Hou, China Dong Song, USA
Guy Carrault, France Seiya Imoto, Japan Xinyuan Song, Hong Kong
Filippo Castiglione, Italy Hsueh-Fen Juan, Taiwan João M. Tavares, Portugal
Carlos Castillo-Chavez, USA Martti Juhola, Finland Jlenia Toppi, Italy
Prem Chapagain, USA Rafik Karaman, Palestine Nelson J. Trujillo-Barreto, UK
Hsiu-Hsi Chen, Taiwan Andrzej Kloczkowski, USA Anna Tsantili-Kakoulidou, Greece
Wai-Ki Ching, Hong Kong Chung-Min Liao, Taiwan Markos G. Tsipouras, Greece
Nadia A. Chuzhanova, UK Ezequiel López-Rubio, Spain Po-Hsiang Tsui, Taiwan
Maria N. D.S. Cordeiro, Portugal Pinyi Lu, USA Gabriel Turinici, France
Cristiana Corsi, Italy Reinoud Maex, Belgium Liangjiang Wang, USA
Irena Cosic, Australia Valeri Makarov, Spain Ruisheng Wang, USA
Qi Dai, China Kostas Marias, Greece David A. Winkler, Australia
Chuangyin Dang, Hong Kong Juan Pablo Martínez, Spain Gabriel Wittum, Germany
Didier Delignières, France Richard J. Maude, Thailand Yu Xue, China
Jun Deng, USA Michele Migliore, Italy Yongqing Yang, China
Thomas Desaive, Belgium John Mitchell, UK Chen Yanover, Israel
David Diller, USA Luminita Moraru, Romania Xiaojun Yao, China
Michel Dojat, France Chee M. Ng, USA Kaan Yetilmezsoy, Turkey
Irini Doytchinova, Bulgaria Michele Nichelatti, Italy Hiro Yoshida, USA
Esmaeil Ebrahimie, Australia Kazuhisa Nishizawa, Japan Henggui Zhang, UK
Georges El Fakhri, USA Francesco Pappalardo, Italy Yuhai Zhao, China
Issam El Naqa, USA Manuel F. G. Penedo, Spain Xiaoqi Zheng, China
Angelo Facchiano, Italy Jesús Picó, Spain
Contents
Mathematical Modeling and Models for Optimal Decision-Making in Health Care
Giedrius Vanagas , Tomas Krilavičius , and Ka Lok Man
Editorial (4 pages), Article ID 2945021, Volume 2019 (2019)
A Simulation Study Comparing Different Statistical Approaches for the Identification of Predictive
Biomarkers
Bernhard Haller , Kurt Ulm, and Alexander Hapfelmeier
Research Article (15 pages), Article ID 7037230, Volume 2019 (2019)
A Data-Driven Hybrid Three-Stage Framework for Hospital Bed Allocation: A Case Study in a Large
Tertiary Hospital in China
Li Luo , Jialing Li , Xueru Xu , Wenwu Shen , and Lin Xiao
An Improved Sliding Window Area Method for T Wave Detection

Haixia Shang, Shoushui Wei , Feifei Liu, Dingwen Wei, Lei Chen, and Chengyu Liu
A Technical Review of Convolutional Neural Network-Based Mammographic Breast Cancer Diagnosis

Lian Zou, Shaode Yu , Tiebao Meng, Zhicheng Zhang, Xiaokun Liang, and Yaoqin Xie
Review Article (16 pages), Article ID 6509357, Volume 2019 (2019)
Automatic Segmentation of Pathological Glomerular Basement Membrane in Transmission Electron

Microscopy Images with Random Forest Stacks
Lei Cao , YanMeng Lu , ChuangQuan Li, and Wei Yang
Minimalistic Approach to Coreference Resolution in Lithuanian Medical Records

Voldemaras Žitkus, Rita Butkienė, Rimantas Butleris, Rytis Maskeliūnas , Robertas Damaševičius ,
and Marcin Woźniak
Brain Tissue Segmentation and Bias Field Correction of MR Image Based on Spatially Coherent FCM
with Nonlocal Constraints
Jianhua Song and Zhe Zhang
Comparing Strategies to Prevent Stroke and Ischemic Heart Disease in the Tunisian Population:
Markov Modeling Approach Using a Comprehensive Sensitivity Analysis Algorithm
Olfa Saidi , Martin O’Flaherty, Nada Zoghlami, Dhafer Malouche, Simon Capewell, Julia A. Critchley,
Piotr Bandosz, Habiba Ben Romdhane, and Maria Guzman Castillo
The Fractional Differential Model of HIV-1 Infection of CD4+ T-Cells with Description of the Effect of
Antiviral Drug Treatment
Bijan Hasani Lichae, Jafar Biazar , and Zainab Ayati
Control of Blood Glucose for Type-1 Diabetes by Using Reinforcement Learning with Feedforward
Algorithm
Phuong D. Ngo , Susan Wei, Anna Holubová, Jan Muzik, and Fred Godtliebsen
Evaluation of the Feasibility of Screening Tau Radiotracers Using an Amyloid Biomathematical

Screening Methodology
Ying-Hwey Nai and Hiroshi Watabe
Continuous Blood Pressure Estimation Based on Two-Domain Fusion Model

Qian Wang, Yajie Xu , Guoqiang Zeng, and Mingshan Sun
Some Similarity Measures of Neutrosophic Sets Based on the Euclidean Distance and Their Application
in Medical Diagnosis
Donghai Liu , Guangyan Liu, and Zaiming Liu
Computational Prediction of the Combined Effect of CRT and LVAD on Cardiac Electromechanical
Delay in LBBB and RBBB
Aulia K. Heikhmakhtiar and Ki M. Lim
Hindawi
Volume 2019, Article ID 2945021, 4 pages
https://doi.org/10.1155/2019/2945021
Editorial
Giedrius Vanagas ,1 Tomas Krilavičius ,2,3 and Ka Lok Man 4,5
1
Institute of Pharmacoeconomics, LT-51052 Kaunas, Lithuania
2
Vytautas Magnus University, Kaunas, Lithuania
3
Lithuania and Baltic Institute of Advanced Technology, Vilnius, Lithuania
4
Xi’an Jiaotong-Liverpool University, Suzhou 215123, China
5
Swinburne University of Technology Sarawak, Kuching, Malaysia
Correspondence should be addressed to Giedrius Vanagas; g.vanagas@gmail.com
Received 31 July 2019; Accepted 1 August 2019; Published 14 August 2019
Copyright © 2019 Giedrius Vanagas et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
Many aspects in the management of healthcare systems are modeling and models for optimal decision-making in health
quantitative, the amount of data within the health care care. We aim to address questions from data analytics, solving
increases by the minute, and, in reality, it makes difficult for problems in predicting outcomes for clinical medicine and
healthcare systems to identify the insights into what is most public health.
valuable for the patients. Data-driven approach to health (or Blood pressure (BP) is one of the indispensable elements
health economic) outcome assessment, artificial intelligence, of physiological health characteristics and a significant in-
and mathematical, computational, methodological, and dicator for predicting and diagnosing hypertension and
technological advances are the core of effective healthcare cardiovascular diseases. Q. Wang et al. proposed a two-
system management [1–3]. domain fusion model to estimate BP continuously from
Modeling in medicine is a valuable tool in the planning pulse wave acquired with a pressure sensor. In more detail,
and evaluation of interventions, especially when a clinical the optimal external pressure applied on the pressure sensor
trial is ethically or logistically impossible [4, 5]. The de- was first determined in order to capture pulse wave in the
velopment of such mathematical models used to simulate radial artery. The captured pulse wave was then processed in
medical outcomes is a growing area in medicine. The both the time and frequency domains via filtering and fast
mathematical modeling is known by various names like Fourier transform. A set of features were extracted from
predictive modeling, simulation, or decision analysis. In these two domains and input into a neural network along
general, modeling techniques are used for health service with blood pressure values measured by a commercial
planning, effectiveness and outcome assessment, healthcare sphygmomanometer for training. Finally, the model was
financing and budget impact assessment, health economic tested on new data for accuracy evaluation, and the proposed
assessments, infectious disease surveillance, health service two-domain fusion method achieved a high degree of ac-
outcomes predicting purposes, and other applications in curacy in measuring blood pressure.
health care. Mathematical modeling is also helpful when H. Shang et al. proposed an improvement for ECG
limitations like a rare event prohibit implementing RCT and analysis, namely, improved sliding window area method for
similar studies or expanding research on actual patients due T wave detection. It allows better detection of T wave onset
to time, ethical, legal, financial, technical, and other limi- and offset, which allows improving clinical diagnosis as well
tations [6, 7]. as daily heart monitoring.
With this special issue, we add to the literature by pro- A. K. Heikhmakhtiar and K. M. Lim proposed com-
viding case studies and practical examples of mathematical putational prediction of the combined effect of CRT and
2 Computational and Mathematical Methods in Medicine
LVAD on cardiac electromechanical delay in the failing SUVR, with respect to binding to a single target, and provides
ventricle with left bundle branch blocked (LBBB) and right some insights to guide the development of in silico models in
bundle branch blocked (RBBB) conditions. The subjects supporting the development of tau radiotracers.
were normal sinus rhythm, LBBB, RBBB, LBBB with CRT- Automatic identification of relevant biomarkers is one of
only, RBBB with CRT-only, LBBB with CRT + LVAD, and the important steps towards personalized treatment. B.
RBBB with CRT + LVAD. The results showed that the CRT- Haller et al. evaluate applicability of a number of methods,
only shortened the total electrical activation time (EAT) in for example, Cox regression with linear interaction, Mul-
the LBBB and RBBB conditions by 20.2% and 17.1%, re- tivariable Fractional Polynomials for Interaction (MFPI),
spectively. The CRT-only reduced the total mechanical ac- Local Partial Likelihood Bootstrap (LPLB), and the Sub-
tivation time (MAT) and electromechanical delay (EMD) of population Treatment Effect Pattern Plot (STEPP), for
the ventricle under LBBB by 21.3% and 10.1%, respectively. biomarker identification. Experiments on randomized
Furthermore, the CRT-only reduced the contractile aden- clinical trials show that the Cox regression works best when
osine triphosphate (ATP) consumption by 5%, increased left interactions are monotonous and the number of events is
ventricular (LV) pressure by 6%, and enhanced cardiac low. When complexity increases, MFPI and LPLB out-
output (CO) by 0.2 L/min under LBBB condition. However, perform other methods. The authors recommend applica-
CRT-only barely affected the ventricle under RBBB condition of statistical methods developed for assessment of
tion. Under the LBBB condition, CRT + LVAD increased LV interactions between continuous biomarkers and treatment
pressure and CO by 10.5% and 0.9 L/min, respectively. instead of arbitrary or data-driven categorization of con-
CRT + LVAD reduced ATP consumption by 15%, shortened tinuous covariates.
the MAT by 23.4%, and shortened the EMD by 15.2%. In The study of D. Liu et al. applies similarity measures of
conclusion, they computationally predicted and quantified single and interval valued neutrosophic sets based on Eu-
that the CRT + LVAD implementation is superior to CRT- clidean distance for diagnostics. Novel theoretical model is
only implementation particularly in HF with LBBB condition. developed in the paper, and its effectiveness is demonstrated
Magnetic resonance imaging has been widely used in on generalized diagnosis, showing that it performs well in
diagnostic imaging for general checkup in clinical practice, solving a multiple criteria decision process. The proposed
especially in detection and diagnosis of brain diseases. similarity measures were applied to medical diagnosis de-
However, brain MR imaging has some lacks such as noise, cision problems, and a number of examples were used to
intensity inhomogeneity, low contrast, and partial volume illustrate the feasibility and effectiveness of the proposed
effect , which brings serious obstacles to segment the brain similarity measure.
MR images. The study of J. Song and Z. Zhang presented a Coreference resolution is a challenging part of natural
novel and more robust method to noise in the brain language processing (NLP) with applications in machine
magnetic resonance imaging, together with a more effective translation, semantic search, and other information retrieval
estimation method of the bias field. and decision support systems. V. Žitkus et al. presented a
Automatic segmentation of different images is one of the method for coreference resolution in the Lithuanian lan-
most important topics in medicine. L. Cao et al. discuss guage and its application for processing e-health records
application of Random Forests Stacks for automatic seg- from a hospital reception. The novelty of their proposed
mentation of pathological glomerular basement membranes method is the ability to process coreferences with minimal
in TEM images. It allows faster observation of morpho- linguistic resources, which is important in linguistic appli-
logical changes, reducing manual and laborious work of cations for rare and endangered languages. Their experi-
specialists. Another exercise in image processing is discussed mental results have shown that coreference resolution is
in J. Song and Z. Zhang. Improvements for brain tissue applicable to the development of NLP-powered online
segmentation and bias field correction of MR images are healthcare services in Lithuania.
proposed and evaluated. The results are promising and Computer-aided models for mammographic breast
potentially can deal with noise in brain MR images. cancer diagnosis (MBCD) have been explored for over thirty
In conventional radiotracer and drug development, poor years [9]. The study of L. Zou et al. dedicated to the tech-
bench-to-bedside translation often results due to the dif- nique of CNN applied in a specific application of MBCD,
ferences between in vitro and in vivo conditions [8]. The and it aims to provide clues on how to use CNN in intelligent
study by Y.-H. Nai and H. Watabe evaluated the feasibility of diagnosis. This study is restricted to peer-reviewed journal
extending the amyloid-validated screening methodology to publications, and consequently, technical details and pros
support the development of tau PET radiotracers, where and cons of each model can be delivered. Furthermore,
more challenges like off-target binding exist. This is the first based on how to use CNN techniques, the MBCD models are
in silico method investigated, which uses the physico- broadly categorized into three groups. One is to design
chemical and pharmacological properties of the compounds shallow models or to modify existing models for decreased
to support tau PET radiotracers developments. 22 PET ra- time cost and medical instances for training; another is to
diotracers reported to bind to tau proteins were investigated, make the best use of a pretrained CNN model by transfer
including 9 clinically applied and tau-focused radiotracers. The learning and parameter fine-tuning; and the third is to take
study supported the use of the screening methodology in ra- advantage of CNN models for feature extraction, while the
diotracer development by allowing comparison of candidate differentiation between malignant and benign lesions is
radiotracers with clinically applied radiotracers based on based on machine learning classifiers. At last, findings,
Computational and Mathematical Methods in Medicine 3
challenges, and limitations are summarized, and some clues beds. In the second stage were used a fitted arrival distri-
on the future work are also given. At present, the design and bution and median LOS as the input to a generic simulation
use of CNN-based MBCD is at its early stage and result- model. In the third stage was built a mixed integer pro-
oriented. The ultimate goal of using deep learning tools is to gramming model using the results obtained in the first two
facilitate clinical practice. This review provides benefits to stages to generate the optimal bed allocation strategy for
scientific researchers, industrial engineers, and those who different departments. The case study demonstrated the
are devoted to intelligent cancer diagnosis. effectiveness of the proposed data-driven hybrid three-stage
The past application of mathematical models in medi- framework and provides hospital bed policy makers with a
cine also has been proven useful in cardiovascular diseases feasible solution for bed allocation.
(CVDs). The study of O. Saidi et al. aimed to describe a Mathematical models are often used and prove their
comprehensive Markov model based on both a probabilistic applicability for optimal decision-making. They are also
multivariate approach and simple linear regression meta- useful to derive estimates of rare or future events from
modeling using the model to evaluate the effects of increases recorded intermediate points. When developing models,
in uptake of stroke treatments, lifestyle changes, and primary decisions are needed about the appropriate level of com-
prevention among the Tunisian population aged 35–94 years plexity to be represented and about model structure and
in 2025. It examined three interventions: improved medical assumptions.
treatments in the acute phase, secondary prevention of
stroke by increasing the prescribing of statins, and primary Conflicts of Interest
prevention aiming to reduce salt intake.
Type-1 diabetes is a condition caused by the lack of The editors declare no conflicts of interest.
insulin hormone, which leads to an excessive increase in
blood glucose level. The glucose kinetics process is difficult to Acknowledgments
control due to its complex and nonlinear nature and with
state variables that are difficult to measure. P. D. Ngo et al. We would like to express our gratefulness to Prof. Katarzyna
proposed a method for automatically calculating the basal Kolasa for participation in the preparation of the call ma-
and bolus insulin doses for patients with type-1 diabetes terials and participation in guest editing some manuscripts.
using reinforcement learning with a feedforward controller. Also, we would like to thank all researchers who submitted
The proposed controller also improved the blood glucose their research work and reviewers who made very important
responses and prevented hypoglycemia condition. Simula- comments and suggestions for authors to improve their
tion of the control system in different uncertain conditions manuscripts.
provided insights on how the inaccuracies of carbohydrate
counting and meal-time reporting affect the performance of Giedrius Vanagas
the control system. As a conclusion, the proposed controller Tomas Krilavičius
is an effective tool for reducing postmeal blood glucose rise Ka Lok Man
and for countering the effects of external known events such
as meal intake and maintaining blood glucose at a healthy References
level under uncertainties.
In the paper of B. H. Lichae et al., the fractional-order [1] A. X. Costa, S. A. Ridley, A. K. Shahani, P. R. Harper,
differential model of HIV-1 infection of CD4+ T-cells with V. De Senna, and M. S. Nielsen, “Mathematical modelling and
the effect of drug therapy has been introduced. There are simulation for planning critical care capacity,” Anaesthesia,
three components: uninfected CD4+ T-cells, x, infected vol. 58, no. 4, pp. 320–327, 2003.
CD4+ T-cells, y, and density of virions in plasma, z. The aim [2] R. B. Chambers, “The role of mathematical modeling in
medical research: “research without patients?”,” The Ochsner
is to gain numerical solution of this fractional-order HIV-1
Journal, vol. 2, no. 4, pp. 218–223, 2000.
model by the Laplace Adomian decomposition method [3] M. Calder, C. Craig, D. Culley et al., “Computational modelling
(LADM). The solution of the proposed model has been for decision-making: where, why, what, who and how,” Royal
achieved in a series form. Moreover, to illustrate the ability Society Open Science, vol. 5, no. 6, article 172096, 2018.
and efficiency of the proposed approach, the solution has [4] G. P. Garnett, S. Cousens, T. B. Hallett, R. Steketee, and
been compared with the solutions of some other numerical N. Walker, “Mathematical models in the evaluation of health
methods. The Caputo sense has been used for fractional programmes,” The Lancet, vol. 378, no. 9790, pp. 515–525, 2011.
derivatives. [5] W. Crown, N. Buyukkaramikli, M. Y. Sir et al., “Application of
Beds are key, scarce medical resources in hospitals. The constrained optimization methods in health services research:
study of L. Luo et al. aimed to balance the utilization of report 2 of the ISPOR Optimization Methods Emerging Good
existing beds in a large tertiary hospital in China. The author Practices Task Force,” Value in Health, vol. 21, no. 9,
pp. 1019–1028, 2018.
developed a data-driven hybrid three-stage framework in-
[6] Y. Zheng, F. Pan, and S. Sorensen, “Modeling treatment se-
corporating data analysis, simulation, and mixed integer quences in pharmacoeconomic models,” PharmacoEconomics,
programming to minimize the gaps in bed occupancy rates vol. 35, no. 1, pp. 15–24, 2017.
(BOR) among different departments. The first stage is to [7] L. Echazu and D. Nocetti, “Priority setting in health care:
calculate the length of stay (LOS) and BOR of each de- disentangling risk aversion from inequality aversion,” Health
partment and identify the departments that need to allocate Economics, vol. 22, no. 6, pp. 730–740, 2013.
[8] Y.-H. Nai, M. Shidahara, C. Seki, and H. Watabe, “Bio-

mathematical screening of amyloid radiotracers with clinical
usefulness index,” Alzheimer’s & Dementia: Translational Re-
search & Clinical Interventions, vol. 3, no. 4, pp. 542–552, 2017.
[9] N. I. R. Yassin, S. Omran, E. M. F. El Houby, and H. Allam,
“Machine learning techniques for breast cancer computer
aided diagnosis using different image modalities: a systematic
review,” Computer Methods and Programs in Biomedicine,
vol. 156, pp. 25–45, 2018.
Hindawi
https://doi.org/10.1155/2019/7037230
Research Article
A Simulation Study Comparing Different Statistical
Approaches for the Identification of Predictive Biomarkers
Bernhard Haller , Kurt Ulm, and Alexander Hapfelmeier

Technical University of Munich, School of Medicine, Institute of Medical Informatics, Statistics and Epidemiology,
Ismaninger Str. 22, 81675 Munich, Germany
Correspondence should be addressed to Bernhard Haller; bernhard.haller@tum.de
Received 1 February 2019; Accepted 22 May 2019; Published 13 June 2019
Guest Editor: Tomas Krilavičius
Copyright © 2019 Bernhard Haller et al. This is an open access article distributed under the Creative Commons Attribution
properly cited.
Identification of relevant biomarkers that are associated with a treatment effect is one requirement for adequate treatment
stratification and consequently to improve health care by administering the best available treatment to an individual patient.
Various statistical approaches were proposed that allow assessing the interaction between a continuous covariate and treatment.
Nevertheless, categorization of a continuous covariate, e.g., by splitting the data at the observed median value, appears to be very
prevalent in practice. In this article, we present a simulation study considering data as observed in a randomized clinical trial with
a time-to-event outcome performed to compare properties of such approaches, namely, Cox regression with linear interaction,
Multivariable Fractional Polynomials for Interaction (MFPI), Local Partial-Likelihood Bootstrap (LPLB), and the Subpopulation
Treatment Effect Pattern Plot (STEPP) method, and of strategies based on categorization of continuous covariates (splitting the
covariate at the median, splitting at quartiles, and using an “optimal” split by maximizing a corresponding test statistic). In
different scenarios with no interactions, linear interactions or nonlinear interactions, type I error probability and the power for
detection of a true covariate-treatment interaction were estimated. The Cox regression approach was more efficient than the other
methods for scenarios with monotonous interactions, especially when the number of observed events was small to moderate.
When patterns of the biomarker-treatment interaction effect were more complex, MFPI and LPLB performed well compared to
the other approaches. Categorization of data generally led to a loss of power, but for very complex patterns, splitting the data into
multiple categories might help to explore the nature of the interaction effect. Consequently, we recommend application of
statistical methods developed for assessment of interactions between continuous biomarkers and treatment instead of arbitrary or
data-driven categorization of continuous covariates.
1. Introduction growth factor receptor (EGFR) mutation in non-small cell

lung cancers (NSCLC) [11]. Consequently, the identification
For medical decision making, predictive biomarkers play an of biomarkers that allow prediction of the treatment effect
important role for various diseases [1–4]. A biomarker is when different treatment options are available is essential to
called “predictive,” if the difference between the effectiveness increase clinical decision making in the sense of a stratified
of two or more treatment options depends on the value of or personalized medicine [12].
that biomarker [5, 6]. In the presence of a qualitative In practice, investigation of such treatment effect het-
biomarker-treatment interaction [7], i.e., when the choice of erogeneity over the range of a certain biomarker in data
the “optimal” treatment for a given patient depends on the obtained from a randomized clinical trial is often performed
patient’s value of a certain biomarker, the biomarker can be by subgroup analyses [13], where the difference in outcome
used for treatment stratification [8]. Biomarkers used in between the study groups, quantified, e.g., by a hazard
clinical practice for treatment stratification are, e.g., the ratio, an odds ratio, or a mean difference, is estimated for
human epidermal growth factor receptor 2 (HER-2) status patient subgroups with similar characteristics [14] and
for breast cancer patients [9, 10] or presence of epidermal compared using a statistical test for interaction, which can be
performed by including the product of the biomarker and articles and further articles including more detailed de-
the variable indicating treatment allocation in an appro- scriptions of the considered methods are given. The setting
priate regression model [15, 16]. While this procedure is of the simulation study and the relevant aspects that were
intuitive and straightforward for categorical variables, e.g., varied are described in Section 2.2. Results of the simulation
gender or presence of comorbidities as diabetes, in- study, namely, observed type I error probabilities for sce-
vestigation of treatment effect heterogeneity with respect to narios with no true biomarker-treatment interactions and
continuous variables, e.g., age or continuously measured estimates for statistical power for scenarios with truly
blood parameters, requires categorization of the variable, present biomarker-treatment interactions, are presented in
when subgroup analyses are to be performed. Such cate- Section 3. A discussion of the results with concluding re-
gorization of continuous variables was criticized due to loss marks and strengths and limitations of our simulation study
of information leading to a loss of power for detection of true is given in Section 4.
interactions, implication of biological implausible effects,
and lack of comparability of results from different studies 2. Methods
[17, 18]. Therefore, various approaches were proposed in the
literature that allow to model and test for treatment effect The methods investigated in the simulation study are de-
heterogeneity over the range of a continuous variable that do scribed in Section 2.1. Details on the settings used in the
not require categorization of the variable [19–21]. simulation study and the data generating process are given
In this article, we describe a simulation study comparing in Section 2.2. Data were generated and analysed using the
different approaches for detection of an interaction between statistical software R [28]. Cox regression was performed
one (predefined) continuous covariate and treatment. We using the function coxph provided in the R library survival
simulated data as they would be expected to be collected in a [29, 30]. For convenience, the continuous covariate of in-
randomized clinical trial intended to compare efficacy of two terest will be called “biomarker” and denoted as Z
treatment groups or of treatment versus placebo. Conse- throughout the section. Treatment allocation will be rep-
quently, patients are allocated randomly into one of two resented by a binary treatment variable T with T � {0, 1},
treatment arms and the distribution of the variable of in- where T � 1 represents, e.g., an experimental treatment and
terest (often referred to as a biomarker [22]) is expected to be T � 0 a placebo control or standard treatment. As it appears
the same for both treatment groups. As most predictive to be the most relevant effect size in practice, homogeneity of
biomarkers were identified for treatment of cancer [23], a the hazard ratio between the study groups in regard to the
time-to-event outcome is considered, as typically overall biomarker of interest was investigated. For all statistical
survival or progression-free survival is considered as pri- tests, a significance level of α � 5% was used. Exact 95%
mary endpoint in randomized phase III oncological trials confidence intervals for rejection probabilities were
[24]. Results obtained by methods relying on categorization calculated.
of the continuous variable as well as methods that do not use
such categorization were investigated. We considered a
method splitting the continuous biomarker at its median to 2.1. Methods Used to Test for a Biomarker-Treatment
determine two subgroups for further analysis, the use of four Interaction
subgroups determined by splitting data at the quartiles, and 2.1.1. Median Split. In many applications investigating
use of an “optimal” cutoff value found by maximization of treatment-effect heterogeneity in regard to a continuous
the Wald statistics of the interaction term in a Cox re- biomarker, individuals are divided into two subgroups of
gression model. Additionally, we applied the Subpopulation equal size. This is achieved by splitting the data at the median
Treatment Effect Pattern Plot (STEPP) approach that in- of the biomarker Z. This procedure will be denoted as
corporates overlapping subgroups [25], the Cox regression “Median split” in this article. A binary indicator variable that
model [26] assuming a linear covariate-treatment in- is assigned the value of one if the biomarker value is above or
teraction term, the Multivariable Fractional Polynomials for equal to the observed median and zero else is derived. To test
Interaction (MFPI) approach that incorporates nonlinear for biomarker-treatment interaction, a Cox regression
transformations for the interaction term [19], and the Local model with this indicator variable, the binary treatment
Partial-Likelihood Bootstrap (LPLB) that uses local esti- indicator, and their product (the interaction term) is fitted to
mates of the treatment effect at different values of the the data. The p value of the Wald test for the interaction term
variable of interest [27]. Different scenarios with absence was used to decide whether the null hypothesis of no
and presence of biomarker-treatment interactions were biomarker-treatment interaction can be rejected on the
investigated in order to estimate and compare type I error prespecified significance level of α � 5%.
probability and statistical power of the different approaches
under the given scenarios. Sample size and censoring dis-
tribution are varying to investigate the impact of these 2.1.2. Quartile Split. As an alternative approach, individuals
characteristics on the outcome. were divided into four subgroups with splits at the corre-
The article is organized as follows. In Section 2, the sponding quartiles of the biomarker of interest (“Quartile
simulation study is described. The different methods used split”). The categorical variable indicating the corresponding
for identification of a biomarker-treatment interaction are subgroup was used as a dummy coded nominal independent
shortly introduced in Section 2.1, and references to original variable in a Cox regression model. Additionally, the binary
treatment indicator and an interaction term between the implies that the log-hazard ratio between the study groups is
dummy coded categorical variable indicating the biomarker linearly associated with the biomarker value. The main ef-
quartile and treatment were included. A likelihood ratio test fects of the biomarker Z, the treatment group T, and their
with three degrees of freedom provided in the R library car product Z × T were used as independent variables in a Cox
[31] was performed to test for presence of a biomarker- regression model. The p value of the Wald test for the in-
treatment interaction. teraction term was considered to decide on rejection of the
null hypothesis of no biomarker-treatment interaction. This
procedure will be called “Cox model with linear interaction”
2.1.3. “Optimal” Split. For this approach, henceforth called or shortly “Cox (linear Int.)” throughout the article.
“Optimal split”, an “optimal” cutoff value for splitting the
continuous variable into two subgroups was determined in a
first step. Of all possible cutoff values (restricted to a 2.1.6. Multivariable Fractional Polynomials for Interaction
minimum subgroup size of 10% of the overall sample size), (MFPI). To allow for nonlinear interaction terms, Royston
the one leading to the largest value of the Wald statistic for and Sauerbrei proposed the Multivariable Fractional Poly-
the interaction term between the dichotomized biomarker nomials for Interaction (“MFPI”) approach [19], which is
and treatment in a Cox regression model also including the based on the Multivariable Fractional Polynomials (MFP)
corresponding main effects as independent variables was approach presented by Royston and Altman [36]. A non-
used to define the subgroups for assessment of treatment linear transformation is considered for the biomarker of
effect heterogeneity. In a second step, these subgroups were interest, and a model including main effects of treatment and
treated as if they were predefined subgroups, and assessment the transformed biomarker as well as their interaction is
of a biomarker-treatment interaction was performed as compared to a model including only the corresponding main
described for the Median split procedure in Section 2.1.1. effects. In the original publication, a model with two
polynomial transformations p1 and p2 (FP2) out of the set
p ∈ {−2, −1, −0.5, 0, 0.5, 1, 2, 3}, where p � 0 indicates a
2.1.4. Subpopulation Treatment Effect Pattern Plot (STEPP) logarithmic transformation, was described. Identification of
Method. The Subpopulation Treatment Effect Pattern Plot the best transformation was proposed to be determined in
(“STEPP”) method was proposed by Bonetti and Gelber the model without an interaction term by finding the
[25]. In the STEPP procedure, heterogeneity of the treatment combination of transformations providing the highest (log-)
effect over the range of a biomarker of interest is assessed by likelihood value (later called flex1 approach). Based on the
estimating the effect in multiple overlapping subgroups. results of a simulation study [37, 38] considering a con-
Additionally, methods for estimation of simultaneous tinuous outcome, an alternative approach with only one
confidence intervals and for testing the null hypothesis of no polynomial transformation (FP1) and separate de-
biomarker-treatment interaction were developed [25, 32]. termination of the best transformation in the model with
Two different versions, a “tail-oriented” and a “sliding and without interaction (flex3, potentially leading to non-
window” approach, were proposed initially. In our simu- nested models) was recommended. We applied both ap-
lation study, we used the “sliding window” approach, where proaches, the FP2-flex1 and the FP1-flex3 approach, to our
the number of individuals within two consecutive subgroups simulated data. To test for presence of a biomarker-
is held (approximately) constant by adding and eliminating treatment interaction, likelihood ratio tests comparing the
the same number of observations and the number of ob- models with and without interaction terms were performed
servations overlapping between two consecutive subgroups for both strategies.
is chosen a priori. For our analysis, the number of in-
dividuals within each subgroup was chosen to be n/5 and the
number of overlapping individuals to be n/10. So, the 2.1.7. Local Partial-Likelihood Bootstrap (LPLB). Another
subgroup sizes were 50, 100, and 200 for scenarios with 250, method proposed in the literature for modelling nonlinear
500, and 1000 observations, and the number of overlapping interaction effects between a continuous biomarker and
observations was 25, 50, and 100, respectively. This led to a treatment is the Local Partial-Likelihood Estimation pro-
total number of nine subgroups considered irrespective of posed by Fan et al. [21]. Liu et al. developed a bootstrapping
the sample size. A test on homogeneity of the hazard ratio method, called Local Partial-Likelihood Bootstrap (“LPLB”),
over all subgroups was performed to test for a biomarker- that allows to test for the presence of an overall treatment
treatment interaction. A permutation test as recommended effect and to test whether the treatment effect is heteroge-
in [32] was conducted using 500 permutations for each neous over the range of a continuous biomarker [27]. In the
simulated dataset. Further details on the STEPP procedure LPLB approach, linear approximations of the treatment
can be found in [33, 34]. For application of STEPP, the R effect estimate at a given biomarker value are obtained by
library stepp [35] was used. first-order Taylor approximations using weighted data in the
local neighbourhood of the biomarker value of interest. The
proposed bootstrap test makes use of the residual bootstrap
2.1.5. Cox Regression Model with Linear Interaction. To [39]. The obtained local estimates of the log-hazard ratio are
avoid categorization of the continuous biomarker of interest compared to the estimate obtained from a standard Cox
Z, a Cox regression model [26] assuming a linear interaction regression model assuming a constant treatment effect over
between Z and treatment T was considered. This procedure the biomarker range. The maximum observed standardized
difference of the local estimates to the constant log-hazard but the hazard ratio between the treatment groups is the
ratio is considered as test statistic. For our simulation study, same for all biomarker values, so no biomarker-treatment
we used the R library lplb [40] to apply the LPLB procedure. interaction is present (Figure 1(b)).
Local estimates were obtained for every decile of the em-
λ(x ∣ z, T � 0) � 0.5 exp􏼐(2z − 1)2 􏼑,
pirical biomarker distribution. A bandwidth, indicating the (3)
amount of observations in the neighbourhood used for local λ(x ∣ z, T � 1) � exp􏼐(2z − 1)2 􏼑,
estimation, of 0.2 was used and an Epanechnikov kernel was
considered for weighting. Five hundred bootstrap samples leading to a hazard ratio of two for all values of Z.
were drawn for each generated dataset. HR(z) � 2. (4)
2.2. Simulation Settings. Data were generated to mimic data

observed in a randomized clinical trial primarily intended Scenario 3. In the third scenario, a true linear interaction
for comparison of two different treatment options. Conse- (on the log-hazard scale) between the biomarker of interest
quently, simulated individuals were randomly allocated to and treatment is present, leading to a hazard ratio between
one of two treatment groups (T � {0, 1}) with equal prob- the treatment groups of one for a biomarker value of Z � 0
ability for each group. The covariate of interest was ran- and to a hazard ratio of exp(0.75) � 2.12 for a value of Z � 1.
domly generated from a uniform distribution with a λ(x ∣ z, T � 0) � 0.7 exp(0.5z),
minimum value of zero and a maximum value of one. Event
times were drawn from an exponential distribution with the λ(x ∣ z, T � 1) � 0.7 exp(0.5z + 0.75z) � 0.7 exp(1.25z).
individual hazard rate depending on the allocated treatment (5)
group and the drawn covariate value as described in Section
2.2.1. Censoring times were drawn from exponential dis- The hazard ratio increases linearly on a logarithmic scale.
tributions with rates as described in Section 2.2.3. The lower HR(z) � exp(0.75z). (6)
value of the two time variables was allocated as observed
time and an observed event was indicated, if the drawn event The scenario is displayed in Figure 1(c)).
time was smaller than the corresponding censoring time,
and a censored observation was indicated else. Scenario 4. In the fourth scenario, a true qualitative
biomarker-treatment interaction, with a higher risk for an
2.2.1. Functional Form. In order to estimate the type I error event under treatment T � 0 as compared to treatment T �
probability and the statistical power for detection of truly 1 for patients with a small value of Z and a higher risk for
present interaction effects associated with the different ap- an event under T � 1 for individuals with a large value of
proaches, different scenarios were investigated. Overall, six Z, is considered (Figure 1(d)). The hazard ratio is
different functional forms were considered, two without monotonically, but not linearly increasing over the bio-
presence of an interaction effect (Scenarios 1 and 2) and four marker range.
scenarios considering different shapes of interaction terms λ(x ∣ z, T � 0) � 0.9,
(Scenarios 3 to 6). All scenarios are visualized in Figure 1, √� (7)
showing the hazard rates used for simulation of the event λ(x ∣ z, T � 1) � 0.35 exp􏼐1.7 z − 0.2z2 − 0.3z􏼑.
times in dependence of the biomarker value (dashed black The qualitative interaction is indicated by a hazard ratio
and solid grey line and black scale/axis) and the resulting being smaller than one for values of Z < 0.424 and larger
hazard ratios (using a logarithmic scale) between the than one for Z > 0.424.
treatment groups (red line and scale/axis). √�
0.35 exp 1.7 z − 0.2z2 − 0.3z􏼁
HR(z) �
Scenario 1. No associations between treatment and risk for 0.9 (8)
an event and between the biomarker of interest and risk for √� 2
an event are present; the hazard rate for each individual was � 0.389 exp􏼐1.7 z − 0.2z − 0.3z􏼑.
set to 1, irrespective of treatment group and biomarker value
(Figure 1(a)).
Scenario 5. In Scenario 5, the risk for an event is similar
λ(x ∣ z, T � 0) � λ(x ∣ z, T � 1) � 1, (1) under both treatments for most of the individuals, but the
where λ(x) indicates the hazard rate as a function of time. risk increases under treatment T � 1 for large values of Z
Consequently, the hazard ratio between the groups is 1 for all (Figure 1(e)).
covariate values, indicating no biomarker-treatment interaction. λ(x ∣ z, T � 0) � 0.9,
(9)
HR(z) � 1. (2) λ(x ∣ z, T � 1) � 0.9 + 1.75z8 .
Consequently, the hazard ratio is close to one for small

Scenario 2. In the second scenario, the hazard rate depends and moderate values of Z but increases for large values. For
on the value of the biomarker Z for both treatment groups, Z � 1, the hazard ratio reaches a value of 2.94.
3 3
4 4
2.5 3 2.5 3
2 2
2 2
Hazard rate λ
Hazard rate λ
Hazard ratio
Hazard ratio
1.5 1.5
1.5 1 1.5 1
1 0.67 1 0.67
0.5 0.5
0.5 0.5
0.33 0.33
0.25 0.25
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Biomarker B Biomarker B
Hazard ratio Hazard ratio
T=1 T=1
T=0 T=0
(a) (b)
3 3
4 4
2.5 3 2.5 3
2 2
2 2
Hazard rate λ
Hazard rate λ
Hazard ratio
Hazard ratio
1.5 1.5
1.5 1 1.5 1
1 0.67 1 0.67
0.5 0.5
0.5 0.5
0.33 0.33
0.25 0.25
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
T=1 T=1
T=0 T=0
(c) (d)
3 3
4 4
2.5 3 2.5 3
2 2
2 2
Hazard rate λ
Hazard rate λ
Hazard ratio
Hazard ratio
1.5 1.5
1.5 1 1.5 1
1 0.67 1 0.67
0.5 0.5
0.5 0.5
0.33 0.33
0.25 0.25
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
T=1 T=1
T=0 T=0
(e) (f )
Figure 1: Scenarios used in the simulation study for comparison of statistical methods. In scenarios 1 and 2 (a, b), data are generated under
the null hypothesis of no biomarker-treatment interaction. In scenarios 3 to 6, which are illustrated in (c) to (f ), the hazard ratio (illustrated
on a log-scale by the red line) depends on the biomarker value, so biomarker-treatment interactions are present.
1.75z8 Interaction (MFPI) procedure with the FP1-flex3 strategy

HR(z) � 1 + . (10) also provided an increased type I error probability of about
0.9
10%. This was mainly caused by those datasets for which
different polynomial transformations for the biomarker
Scenario 6. In the sixth scenario, the hazard ratio for group were selected for models with and without consideration of a
T � 0 depending on Z follows a U-shape, while the hazard biomarker-treatment interaction, leading to a comparison of
ratio for T � 1 is inversely U-shaped (Figure 1(f )). nonnested regression models. When only simulated datasets
were considered, in which the same transformations were
λ(x ∣ z, T � 0) � 0.75 exp􏼐0.4(2z − 1)2 􏼑,
(11) used for the models with and without interaction term and
λ(x ∣ z, T � 1) � 1.25 exp􏼐−0.5(2z − 1)2 􏼑. consequently two nested models were compared, the esti-
mated type I error probabilities ranged from 3.8% to 6.6%.
This setting leads to a qualitative biomarker-treatment Contrarily, for datasets with different chosen trans-
interaction with lower risks for an event under T � 1 for formations, the null hypothesis was falsely rejected in 14.1%
small and large values of Z and lower risks under T � 0, else to 23.8% of the corresponding simulation runs. For the
indicated by an inversely U-shaped hazard ratio over the simulations under Scenario 2 with a low sample size of 250
range of Z. observations and a high amount of censored observations,
5 leading to an expected number of about 83 events, type I
HR(z) � exp􏼐−0.9(2z − 1)2 􏼑. (12) error frequencies exceeding the nominal significance level
3
were observed for all methods.
In Figures 3 (Scenarios 3 and 4) and 4 (Scenarios 5 and 6)
and in Tables 2 (Scenarios 3 and 4) and 3 (Scenarios 5 and 6),
the results of the scenarios with true biomarker-treatment
2.2.2. Sample Size. In order to evaluate whether properties
interaction are presented. Consequently, the frequency of
of the methods under consideration are related to the sample
rejected null hypotheses can be interpreted as an estimate for
size of the trial, three different settings for sample sizes were
the statistical power of the methods under the corresponding
chosen. The generated datasets included 250, 500, or 1000
settings. As the procedure using two subgroups defined by
individuals, which appear to be typical sample sizes for
an optimal, data-driven cutpoint (Optimal split) and the
randomized clinical trials.
MFPI (FP1-flex3) approach provided type I error proba-
bilities relevantly exceeding the nominal level of α � 5%,
2.2.3. Censoring Distribution. In addition to the sample size, these procedures are not considered in the comparison of
censoring distributions were varied to produce scenarios statistical power and are consequently not displayed in
with different numbers of observed events. Censoring times Figures 3 and 4. Nevertheless, the results are presented in
were drawn from exponential distributions with hazard rates Tables 2 and 3 in italics for completeness.
of λcens � 0.3 or λcens � 2, respectively, to produce scenarios For the scenario fulfilling the assumption of the standard
with censoring proportions of about 25% and about 67%, Cox regression model with a linear interaction term (Sce-
leading to numbers of about 188, 375, and 750 expected nario 3), the Cox model with linear interaction out-
events for scenarios with low amount of censored obser- performed all the other investigated methods by achieving
vations and of 83, 167, and 333 expected events for scenarios the highest observed statistical power (Figures 3(a) and 3(b)
with high amount of censored observations. and Table 2). The MFPI (FP2-flex1) approach performed
slightly better than the approach using two subgroups de-
3. Results fined by a split at the median of the variable when the
number of expected events was large, but for the scenario
For each of the 36 scenarios described in Section 2.2, 1000 with 1000 observations and a low amount of censored
datasets were generated and the methods presented in observations, the observed power was about 10 percentage
Section 2.1 were applied. The p value of the corresponding points lower for these methods as compared to the Cox
statistical test on biomarker-treatment interaction was regression model with an interaction term considering the
saved and compared to the conventional significance level biomarker as continuous variable (Cox model with linear
of α � 5%. Resulting frequencies of type I errors, i.e., pro- interaction: 83.8%; MFPI (FP2-flex1): 74.7%; Median split:
portions of simulated datasets for which a statistically sig- 70.2%). The method splitting the data into four subgroups
nificant biomarker-treatment interaction was found, (Quartile split), the STEPP, and the LPLB performed worse
although it is not present in the corresponding scenario than the other approaches.
(Scenarios 1 and 2), are shown in Figure 2 for all considered In Scenario 4, considering a situation with a slightly
methods and are also tabulated with 95% confidence in- nonlinear interaction, the Cox regression model consid-
tervals in Table 1. It can be seen that for the method using the ering the continuous biomarker performed best again,
“optimal” cutpoint to define two subgroups to be compared, followed (at least for scenarios with a large number of
the probability for a false-positive result was about 50% for events) by the MFPI (FP2-flex1) approach. For small to
both scenarios simulating data under the null hypothesis, moderate event numbers, the methods relying on cate-
irrespective of sample size and amount of censored obser- gorization of the data (Median split and Quartile split)
vations. The Multivariable Fractional Polynomial for performed similarly to MFPI (FP2-flex1). With the chosen
Low censoring High censoring
Median split Median split

Quartile split Quartile split
Optimal split Optimal split
STEPP STEPP
Cox (linear int.) Cox (linear int.)
MFPI (FP1-flex3) MFPI (FP1-flex3)
LPLB LPLB

Scenario 1
Scenario 1
STEPP STEPP
LPLB LPLB

STEPP STEPP
LPLB LPLB
0 10 20 30 40 50 0 10 20 30 40 50
Estimated type I error probability (%) Estimated type I error probability (%)
Sample size Sample size

250 250
500 500
1000 1000
(a) (b)

STEPP STEPP
LPLB LPLB

Scenario 2
Scenario 2
STEPP STEPP
LPLB LPLB

STEPP STEPP
LPLB LPLB
0 10 20 30 40 50 0 10 20 30 40 50
Estimated type I error probability (%) Estimated type I error probability (%)

250 250
500 500
1000 1000
(c) (d)
Figure 2: Results of scenarios simulated under the null hypothesis of no biomarker-treatment interaction. Bars represent relative fre-
quencies of falsely rejected null hypotheses.
Table 1: Estimated type I error probabilities with exact 95% confidence intervals (in brackets) for Scenarios 1 and 2 for all investigated
methods.
n � 250 n � 500 n � 1000
Low cens. High cens. Low cens. High cens. Low cens. High cens.
4.7% 5.3% 4.9% 4.4% 5.0% 6.4%
Median split
(3.5–6.2%) (4.0–6.9%) (3.6–6.4%) (3.2–5.9%) (3.7–6.5%) (5.0–8.1%)
6.3% 5.3% 4.2% 5.1% 3.5% 5.7%
Quartile split
(4.9–8.0%) (4.0–6.9%) (3.0–5.6%) (3.8–6.7%) (2.4–4.8%) (4.3–7.3%)
43.6% 39.9% 45.8% 40.9% 45.5% 46.6%
Optimal split
(40.5–46.7%) (36.8–43.0%) (42.7–48.9%) (37.8–44.0%) (42.4–48.6%) (43.5–49.7%)
4.8% 5.9% 5.1% 4.2% 4.0% 6.3%
STEPP
(3.6–6.3%) (4.5–7.5%) (3.8–6.7%) (3.0–5.6%) (2.9–5.4%) (4.9–8.0%)
Scenario 1
4.9% 5.2% 4.8% 4.4% 4.8% 5.8%
Cox (linear int.)
(3.6–6.4%) (3.9–6.8%) (3.6–6.3%) (3.2–5.9%) (3.6–6.3%) (4.4–7.4%)
10.1% 10.6% 10.4% 12.0% 10.3% 13.5%
MFPI (FP1-flex3)
(8.3–12.1%) (8.8–12.7%) (8.6–12.5%) (10.1–14.2%) (8.5–12.4%) (11.4–15.8%)
4.9% 5.6% 4.9% 4.7% 6.0% 7.0%
MFPI (FP2-flex1)
(3.6–6.4%) (4.3–7.2%) (3.6–6.4%) (3.5–6.2%) (4.6–7.7%) (5.5–8.8%)
4.6% 4.5% 4.2% 4.4% 3.9% 5.8%
LPLB
(3.4–6.1%) (3.3–6.0%) (3.0–5.6%) (3.2–5.9%) (2.8–5.3%) (4.4–7.4%)
4.2% 6.0% 5.7% 4.2% 5.7% 4.3%
Median split
(3.0–5.6%) (4.6–7.7%) (4.3–7.3%) (3.0–5.6%) (4.3–7.3%) (3.1–5.7%)
5.3% 6.7% 5.1% 4.5% 5.0% 5.4%
Quartile split
(4.0–6.9%) (5.2–8.4%) (3.8–6.7%) (3.3–6.0%) (3.7–6.5%) (4.1–7.0%)
50.8% 42.8% 53.9% 45.9% 52.0% 47.4%
Optimal split
(47.7–53.9%) (39.7–45.9%) (50.8–57.0%) (42.8–49.0%) (48.9–55.1%) (44.3–50.5%)
5.4% 8.2% 6.8% 6.8% 7.8% 6.9%
STEPP
(4.1–7.0%) (6.6–10.1%) (5.3–8.5%) (5.3–8.5%) (6.2–9.6%) (5.4–8.7%)
Scenario 2
4.5% 8.2% 4.8% 6.4% 5.0% 5.2%
Cox (linear int.)
(3.3–6.0%) (6.6–10.1%) (3.6–6.3%) (5.0–8.1%) (3.7–6.5%) (3.9–6.8%)
8.1% 12.6% 9.1% 10.3% 8.1% 7.8%
MFPI (FP1-flex3)
(6.5–10.0%) (10.6–14.8%) (7.4–11.1%) (8.5–12.4%) (6.5–10.0%) (6.2–9.6%)
5.5% 6.5% 6.5% 5.6% 4.1% 5.9%
MFPI (FP2-flex1)
(4.2–7.1%) (5.1–8.2%) (5.1–8.2%) (4.3–7.2%) (3.0–5.5%) (4.5–7.5%)
6.2% 5.8% 7.3% 4.8% 7.7% 6.1%
LPLB
(4.8–7.9%) (4.4–7.4%) (5.8–9.1%) (3.6–6.3%) (6.1–9.5%) (4.7–7.8%)
settings, the estimated power for LPLB and STEPP was (87.5%), Cox regression with linear interaction (76.2%),
smaller than for the other investigated methods or LPLB (83.4%).
(Figures 3(c) and 3(d) and Table 2). In Scenario 6, the only investigated scenario with
In the rather complex Scenario 5 with an almost identical nonmonotonous hazard ratio over the range of the bio-
risk for an event under both treatments for most patients marker of interest, the Cox model with linear interaction and
and an increasing difference between treatments for large the procedure defining subgroups at the observed median
values of the biomarker, the MFPI (FP2-flex1) approach (Median split) were not able to identify the present
performed best in scenarios with a large number of observed biomarker-treatment interaction (estimated power between
events. In scenarios with a high amount of censored ob- 4.6% and 6.2% for Cox model with linear interaction and
servations, the Cox model with linear interaction performed between 3.9% and 6.4% for Median split). The highest
slightly better (small to moderate sample size) or very similar empirical power was observed for LPLB and the method
(large sample size) to MFPI (FP2-flex1) (Figures 4(a) and defining four subgroups at the observed quartiles (Quartile
4(b) and Table 3). When censoring was low and sample size split). STEPP and MFPI were able to identify the association
was large, the LPLB approach reached an observed power between biomarker and treatment effect in a relevant
that was close to MFPI (FP2-flex1) and slightly better than amount of generated datasets but performed worse than
the Cox regression model. While categorization using a LPLB and Quartile split (Figures 4(c) and 4(d) and Table 3).
Median split was much worse than the other methods for
most settings under Scenario 5 (e.g., with an observed power 4. Discussion
for n � 1000 and low amount of censored observations of
46.6%), splitting the study population in four subgroups It is well known and accepted that different patients react
(Quartile split) provided results that were relevantly better differently to the same treatment. Consequently, for making
than Median split (estimated power for the mentioned a treatment decision, characteristics of the patient or of the
settings of 70.9% ), but worse than MFPI (FP2-flex1) disease, e.g., of a tumour, should be considered. Predictive

STEPP STEPP
LPLB LPLB

Scenario 3
Scenario 3
STEPP STEPP
LPLB LPLB

STEPP STEPP
LPLB LPLB
0 20 40 60 80 100 0 20 40 60 80 100
Estimated power (%) Estimated power (%)
250 250
500 500
1000 1000
(a) (b)

STEPP STEPP
LPLB LPLB

Scenario 4
Scenario 4
STEPP STEPP
LPLB LPLB

STEPP STEPP
LPLB LPLB
0 20 40 60 80 100 0 20 40 60 80 100
250 250
500 500
1000 1000
(c) (d)
Figure 3: Results of scenarios simulated under the alternative hypothesis of a truly present biomarker-treatment interaction. Bars represent
relative frequencies of correctly rejected null hypotheses.

STEPP STEPP
LPLB LPLB

Scenario 5
Scenario 5
STEPP STEPP
LPLB LPLB

STEPP STEPP
LPLB LPLB
0 20 40 60 80 100 0 20 40 60 80 100
250 250
500 500
1000 1000
(a) (b)

STEPP STEPP
LPLB LPLB

Scenario 6
Scenario 6
STEPP STEPP
LPLB LPLB

STEPP STEPP
LPLB LPLB
0 20 40 60 80 100 0 20 40 60 80 100
250 250
500 500
1000 1000
(c) (d)
Figure 4: Results of scenarios simulated under the alternative hypothesis of a truly present biomarker-treatment interaction. Bars
represented relative frequencies of correctly rejected null hypotheses.
Table 2: Estimated power with exact 95% confidence intervals (in brackets) for Scenarios 3 and 4 for all investigated methods.
n � 250 n � 500 n � 1000
23.5% 14.1% 40.8% 25.1% 70.2% 40.9%
Median split
(20.9–26.3%) (12.0–16.4%) (37.7–43.9%) (22.4–27.9%) (67.3–73.0%) (37.8–44.0%)
19.2% 12.4% 36.4% 17.7% 66.0% 31.2%
Quartile split
(16.8–21.8%) (10.4–14.6%) (33.4–39.5%) (15.4–20.2%) (63.0–68.9%) (28.3–34.2%)
71.4% 57.6% 86.4% 71.7% 97.1% 84.6%
Optimal split
(68.5–74.2%) (54.5–60.7%) (84.1–88.5%) (68.8–74.5%) (95.9–98.0%) (82.2–86.8%)
13.4% 10.5% 29.2% 15.5% 55.1% 26.4%
STEPP
(11.3–15.7%) (8.7–12.6%) (26.4–32.1%) (13.3–17.9%) (52.0–58.2%) (23.7–29.2%)
Scenario 3
29.9% 15.2% 54.2% 31.6% 83.8% 51.9%
Cox (linear int.)
(27.1–32.8%) (13–17.6%) (51.1–57.3%) (28.7–34.6%) (81.4–86.0%) (48.8–55.0%)
30.2% 18.2% 54.2% 32.5% 82.8% 51.1%
MFPI (FP1-flex3)
(27.4–33.2%) (15.9–20.7%) (51.1–57.3%) (29.6–35.5%) (80.3–85.1%) (48–54.2%)
20.4% 12.7% 42.9% 21.0% 74.7% 41.2%
MFPI (FP2-flex1)
(17.9–23.0%) (10.7–14.9%) (39.8–46.0%) (18.5–23.7%) (71.9–77.4%) (38.1–44.3%)
15.2% 9.1% 32.7% 16.2% 61.2% 30.9%
LPLB
(13.0–17.6%) (7.4–11.1%) (29.8–35.7%) (14.0–18.6%) (58.1–64.2%) (28.0–33.9%)
26.8% 14.1% 52.3% 24.4% 78.7% 38.7%
Median split
(24.1–29.7%) (12.0–16.4%) (49.2–55.4%) (21.8–27.2%) (76–81.2%) (35.7–41.8%)
22.8% 14.2% 48.6% 24.2% 79.4% 37.6%
Quartile split
(20.2–25.5%) (12.1–16.5%) (45.5–51.7%) (21.6–27.0%) (76.8–81.9%) (34.6–40.7%)
77.6% 57.8% 92.3% 72.6% 99.1% 88.8%
Optimal split
(74.9–80.1%) (54.7–60.9%) (90.5–93.9%) (69.7–75.3%) (98.3–99.6%) (86.7–90.7%)
14.8% 8.3% 36.8% 17.6% 68.9% 29.9%
STEPP
(12.7–17.2%) (6.7–10.2%) (33.8–39.9%) (15.3–20.1%) (65.9–71.8%) (27.1–32.8%)
Scenario 4
33.9% 17.7% 64.4% 32.4% 89.8% 49.2%
Cox (linear int.)
(31.0–36.9%) (15.4–20.2%) (61.3–67.4%) (29.5–35.4%) (87.8–91.6%) (46.1–52.3%)
40.4% 25.8% 72.0% 38.8% 92.4% 56.3%
MFPI (FP1-flex3)
(37.3–43.5%) (23.1–28.6%) (69.1–74.8%) (35.8–41.9%) (90.6–94.0%) (53.2–59.4%)
22.3% 14.5% 51.2% 25.6% 84.7% 41.3%
MFPI (FP2-flex1)
(19.8–25.0%) (12.4–16.8%) (48.1–54.3%) (22.9–28.4%) (82.3–86.9%) (38.2–44.4%)
19.4% 8.4% 44.1% 20.1% 75.2% 33.8%
LPLB
(17.0–22.0%) (6.8–10.3%) (41.0–47.2%) (17.7–22.7%) (72.4–77.8%) (30.9–36.8%)
Due to increased type I error probabilities, results for Optimal split and MFPI (FP1-flex3) are presented in italics.
biomarkers, i.e., variables that are associated with the biomarker and the treatment indicator as independent
treatment effect, e.g., a hazard ratio between two treatment variable in a regression model assuming a linear interaction
groups, play an important role for treatment selection. term. To allow a more flexible modelling of relationships
Evidence, whether a biomarker is truly predictive, can only between treatment effects and biomarker values, various
be derived from randomized trials involving patients with methods relaxing the linearity assumption for the in-
different values of the biomarker of interest [8]. In practice, teraction term, e.g., the Subpopulation Treatment Effect
treatment effect heterogeneity over different factors of a Pattern Plot (STEPP), the Multivariable Fractional Poly-
categorical variable or over the range of a continuous var- nomials for Interaction (MFPI) [19], or the Local Partial-
iable in data collected in a randomized clinical trial is often Likelihood Bootstrap (LPLB) [27] approach, were
analysed by the means of subgroup analyses, estimating the developed.
treatment effect within patients with similar characteristics Comparisons between those methods rarely exist in the
and comparing treatment effects between subgroups using a literature. Royston and Sauerbrei applied the MFPI and the
test on interaction [14]. While this procedure is straight- STEPP method to different datasets [44]. Recently, we in-
forward for categorical variables, it relies on categorization vestigated the interaction between age and treatment in a
of continuous variables. It was shown for different research randomized trial comparing carotid artery stenting (CAS) to
questions that categorization leads to a loss of power for carotid endarterectomy (CEA) for patients with symp-
detection of true associations [41, 42], and the interpretation tomatic, severe carotid artery stenosis (SPACE trial [45, 46]).
of subgroup analyses based on categorized continuous In this analysis, very similar results were obtained from
variables was often criticized due to its lack of biological different methods including Cox regression with linear
plausibility and its increased chance of spurious findings interaction, STEPP, MFPI, and LPLB [47]. To our best
[17, 18, 43]. One common approach to investigate such knowledge, only a small number of simulation studies were
interactions between continuous biomarkers and treatment performed to compare the properties of the different pro-
without categorization is the inclusion of the product of the cedures under known scenarios. Royston and Sauerbrei
Table 3: Estimated power with exact 95% confidence intervals (in brackets) for Scenarios 5 and 6 for all investigated methods.
n � 250 n � 500 n � 1000
16.1% 10.4% 26.7% 17.7% 46.6% 29.3%
Median split
(13.9–18.5%) (8.6–12.5%) (24.0–29.6%) (15.4–20.2%) (43.5–49.7%) (26.5–32.2%)
23.0% 10.5% 41.3% 21.2% 70.9% 40.1%
Quartile split
(20.4–25.7%) (8.7–12.6%) (38.2–44.4%) (18.7–23.9%) (68.0–73.7%) (37.0–43.2%)
79.7% 60.2% 93.8% 78.3% 99.4% 91.9%
Optimal split
(77.1–82.2%) (57.1–63.2%) (92.1–95.2%) (75.6–80.8%) (98.7–99.8%) (90.0–93.5%)
15.9% 7.4% 35.0% 14.8% 68.7% 31.0%
STEPP
(13.7–18.3%) (5.9–9.2%) (32.0–38.0%) (12.7–17.2%) (65.7–71.6%) (28.1–34.0%)
Scenario 5
25.6% 16.3% 47.0% 30.6% 76.2% 51.1%
Cox (linear int.)
(22.9–28.4%) (14.1–18.7%) (43.9–50.1%) (27.8–33.6%) (73.4–78.8%) (48.0–54.2%)
39.7% 22.6% 67.3% 39.7% 93.8% 66.5%
MFPI (FP1-flex3)
(36.7–42.8%) (20.0–25.3%) (64.3–70.2%) (36.7–42.8%) (92.1–95.2%) (63.5–69.4%)
26.8% 13.2% 50.6% 27.1% 87.5% 51.8%
MFPI (FP2-flex1)
(24.1–29.7%) (11.2–15.5%) (47.5–53.7%) (24.4–30.0%) (85.3–89.5%) (48.7–54.9%)
22.5% 9.3% 46.9% 21.9% 83.4% 45.3%
LPLB
(19.9–25.2%) (7.6–11.3%) (43.8–50.0%) (19.4–24.6%) (80.9–85.7%) (42.2–48.4%)
5.3% 3.9% 5.1% 5.1% 6.4% 4.9%
Median split
(4.0–6.9%) (2.8–5.3%) (3.8–6.7%) (3.8–6.7%) (5.0–8.1%) (3.6–6.4%)
24.3% 13.0% 42.0% 16.0% 73.8% 36.5%
Quartile split
(21.7–27.1%) (11.0–15.2%) (38.9–45.1%) (13.8–18.4%) (71.0–76.5%) (33.5–39.6%)
73.8% 56.5% 88.1% 67.1% 97.6% 86.3%
Optimal split
(71.0–76.5%) (53.4–59.6%) (85.9–90.0%) (64.1–70.0%) (96.4–98.5%) (84.0–88.4%)
14.8% 7.9% 31.4% 12.7% 61.6% 25.9%
STEPP
(12.7–17.2%) (6.3–9.7%) (28.5–34.4%) (10.7–14.9%) (58.5–64.6%) (23.2–28.7%)
Scenario 6
6.1% 5.9% 4.8% 5.1% 6.2% 4.6%
Cox (linear int.)
(4.7–7.8%) (4.5–7.5%) (3.6–6.3%) (3.8–6.7%) (4.8–7.9%) (3.4–6.1%)
23.0% 16.5% 30.6% 18.5% 50.9% 27.3%
MFPI (FP1-flex3)
(20.4–25.7%) (14.3–18.9%) (27.8–33.6%) (16.1–21.0%) (47.8–54.0%) (24.6–30.2%)
20.0% 11.2% 28.6% 12.8% 45.9% 26.9%
MFPI (FP2-flex1)
(17.6–22.6%) (9.3–13.3%) (25.8–31.5%) (10.8–15.0%) (42.8–49.0%) (24.2–29.8%)
22.2% 10.4% 41.5% 17.9% 74.6% 36.7%
LPLB
(19.7–24.9%) (8.6–12.5%) (38.4–44.6%) (15.6–20.4%) (71.8–77.3%) (33.7–39.8%)
Due to increased type I error probabilities, results for Optimal split and MFPI (FP1-flex3) are presented in italics.
performed a simulation study to compare different MFPI not involved in the development or publication of any of
strategies to other regression models and approaches relying them.
on categorization of continuous variables in settings with a As to be expected, we observed that the procedure using
continuous outcome [37, 38]. Under all the different MFPI an optimal cutoff value determined by maximizing the Wald
strategies investigated there, the MFPI (FP1-flex3) approach, statistic of the interaction term between the dichotomized
using one polynomial transformation and allowing for biomarker of interest and treatment in a Cox regression
different functional forms in the models with and without model for definition of the subgroups leads to a tremen-
considering a covariate-treatment interaction, was identified dously increased type I error probability of about 50%. This
as the “best” MFPI approach. Bonetti et al. performed a was observed similarly in simulations presented by Altman
simulation study to evaluate the impact of the parameter et al. who investigated the naı̈ve use of minimum p value
settings of the STEPP approach on type I error and statistical categorization of a potentially prognostic variable [49].
power and compared the results to those of a Cox regression Interestingly, an increased type I error probability of about
model with linear interaction term [32]. Liu et al. also 10% in both scenarios with data simulated under the null
compared performance of their proposed LPLB approach to hypothesis was also observed for the MFPI (FP1-flex3)
the Cox regression model with a linear interaction term [27]. approach irrespective of sample size and censoring distri-
Due to the lack of information on the properties of different bution. This was caused by datasets for which different
available methods proposed in the literature for identifi- transformations were selected for the models with and
cation of a biomarker-treatment interaction, we performed a without an interaction term. In the simulation study by
simulation study comparing estimates for type I error Royston and Sauerbrei [37], no relevant increase in the
probability and statistical power of relevant methods under probability of false-positive findings was identified for the
various scenarios. Our aim was to perform a study in the MFPI (FP1-flex3) approach for most of their investigated
sense of a “neutral” simulation study as described in [48] as scenarios with observed relative frequencies of type one
we do not favour any of the investigated methods and were errors ranging from 5% to 7%. Only for scenarios with
complex functional forms and a covariate of interest fol- overlapping individuals in STEPP or the number of points
lowing a skewed distribution (called “badly behaved dis- used for local estimation and the bandwidth in LPLB. As
tribution of x” in [37]), an increased type I error probability we only used one setting for each of the methods as de-
of up to 20% was found. Maybe this problem is less pro- scribed in Section 2.1, our findings are only valid for these
nounced in a linear regression setting with quantitative specific choices, but might not transfer to the methods in
outcome than for our investigated time-to-event endpoint. general. Further simulation studies are needed to in-
The originally proposed MFPI (FP2-flex1) approach did not vestigate the role of the different tuning parameters on the
lead to an increased probability of false-positive results and performance of these methods. In practical applications,
performed generally well for all scenarios. While it was subject knowledge could allow more adequate specifica-
superior to all other methods in a scenario with a hazard tions, which might improve performance of the methods
ratio constant over a wide range of the biomarker and in- compared to our fixed settings. Additionally, we only
creasing for individuals with large values when the number investigated one potential predictive biomarker and
of events was large, it was slightly less efficient than a Cox treated it as if investigation of interaction of that bio-
regression model with a linear interaction term in the marker with treatment was the prespecified primary re-
presence of a truly linear or close to linear biomarker- search question. In practice, these kinds of analyses will
treatment interaction. Generally, the Cox regression often be performed as exploratory secondary or add-on
model with a linear interaction term performed better than analyses, potentially involving multiple biomarkers of
the other investigated methods for many scenarios. It interest, and multiplicity issues typically evolving in these
provided an acceptable probability of false-positive results situations will have to be addressed adequately. If testing
and higher statistical power than all other methods in the the interaction between a predefined biomarker and
scenario with a truly linear interaction. For small to mod- treatment is of major interest, this has to be considered in
erate event numbers, the Cox regression model also out- the planning phase of a clinical trial and consequently in
performed the other methods in scenarios with nonlinear the sample size calculation, as often a large sample size is
monotonous interaction effects. In one scenario with data necessary to detect biomarker-treatment interactions [50].
generated to provide a nonmonotonous interaction effect It has to be considered that our simulation study only
over the range of the biomarker of interest, the Cox re- aims at detection of biomarker-treatment interactions.
gression model assuming a linear interaction term was not According to Chen et al., three steps are needed to establish a
able to detect this association. For the LPLB procedure, type predictive biomarker in clinical practice: identification of a
I error frequencies did not exceed the nominal significance biomarker, selection of adequate subgroups for treatment
level relevantly and adequate statistical power as compared stratification, and assessment of clinical utility. Conse-
to the other methods was observed for scenarios with quently, after identification of a predictive biomarker,
complex functional form of the hazard ratio over the bio- subgroups that should be treated by different treatment
marker range. The procedure splitting the data into two options have to be identified. For continuous biomarkers,
subgroups (Median split) led to decreased power for most this could be achieved by either application of classification
scenarios, which was also described for other research techniques [51] or by exploring the pattern of the treatment
questions dealing with categorization of continuous cova- effect estimate over the range of the biomarker value. In-
riates [41, 42]. For complex associations, the split into a small tuitive visualization as provided by STEPP or by the
number of subgroups might be an adequate first step for data “treatment effect plot” [52] of the MFPI procedure can be
exploration, which was also recommended in the EMA helpful. Additionally, further aspects such as potential risks,
guideline on subgroup analyses [14], or might be used for patient acceptance, and costs have to be taken into account.
verification of nonlinear associations found by a corre- Clinical utility might be investigated by randomized clinical
sponding method as also recommended in [38]. trials using biomarker-stratified or biomarker-strategy de-
Our simulation study has several limitations. Due to signs as described by Ondra et al. [53].
limited time and space, only a small number of different As a conclusion of our simulation study, we recommend
scenarios could be investigated. We considered two sce- to perform more detailed and sophisticated analyses for
narios in which data were generated under the null hy- detection of biomarker-treatment interactions than the
pothesis of no biomarker-treatment interaction and four commonly performed subgroup analyses involving di-
settings with true biomarker-treatment interactions of chotomization of continuous variables. Cox regression
different shapes. Additionally, we varied the sample size models considering linear interaction terms will increase the
and used two different amounts of censored observations. probability for detection of true interactions as compared to
We did not vary further aspects of the data generating the use of dichotomized variables in many applications.
process as the distribution of the covariate of interest or Methods developed for detection of nonlinear interactions
the influence of further covariates. While some of the can help to identify predictive biomarkers in the presence of
methods as fitting a Cox regression model with linear complex patterns. We believe that better use of available
interaction to the data or application of the MFPI ap- statistical methods will help to identify and establish pre-
proach do not rely on the specification of tuning pa- dictive biomarkers and increase the number, up to now
rameters, other methods such as STEPP or LPLB allow a limited [54], of biomarkers used in clinical practice for
greater level of user involvement by letting the applicant treatment stratification and consequently help to improve
choose, e.g., the size of the subgroups or the number of health care for individual patients.
Data Availability [13] R. Wang, S. W. Lagakos, J. H. Ware, D. J. Hunter, and

J. M. Drazen, “Statistics in medicine-reporting of subgroup
All findings are based on simulated data. R code for data analyses in clinical trials,” New England Journal of Medicine,
generation can be obtained from the corresponding author vol. 357, no. 21, pp. 2189–2194, 2007.
upon reasonable request. [14] EMA/CHMP, Guideline on the Investigation of Subgroups in
Confirmatory Trial (Draft), EMA/CHMP/539146/2013,
Committee for Medicinal Products for Human Use, London,
Conflicts of Interest UK, 2014.
[15] W. Vach, Regression Models as a Tool in Medical Research,
The authors declare that there are no conflicts of interest Boca Raton, FL, USA, CRC Press, 2012.
regarding the publication of this article. [16] P. Royston and W. Sauerbrei, Multivariable Model-Building:
A Pragmatic Approach to Regression Analysis Based on
Fractional Polynomials for Modelling Continuous Variables,
Acknowledgments John Wiley & Sons, Chichester, UK, 2008.
[17] S. F. Assmann, S. J. Pocock, L. E. Enos, and L. E. Kasten,
This work was supported by the German Research Foun- “Subgroup analysis and other (mis)uses of baseline data in
dation (DFG) and the Technical University of Munich clinical trials,” The Lancet, vol. 355, no. 9209, pp. 1064–1069,
within the funding programme Open Access Publishing. 2000.
[18] S. J. Pocock, S. E. Assmann, L. E. Enos, and L. E. Kasten,
“Subgroup analysis, covariate adjustment and baseline
References comparisons in clinical trial reporting: current practice and
[1] E. Thunnissen, K. van der Oord, and M. Den Bakker, problems,” Statistics in Medicine, vol. 21, no. 19, pp. 2917–
“Prognostic and predictive biomarkers in lung cancer. a re- 2930, 2002.
view,” Virchows Archiv, vol. 464, no. 3, pp. 347–358, 2014. [19] P. Royston and W. Sauerbrei, “A new approach to modelling
[2] W. H. Robinson and R. Mao, “Biomarkers to guide clinical interactions between treatment and continuous covariates in
therapeutics in rheumatology?,” Current Opinion in Rheu- clinical trials by using fractional polynomials,” Statistics in
matology, vol. 28, no. 2, pp. 168–175, 2016. Medicine, vol. 23, no. 16, pp. 2509–2525, 2004.
[3] R. Hoefflin, A.-L. Geißler, R. Fritsch et al., “Personalized [20] L. Tian, A. A. Alizadeh, A. J. Gentles, and R. Tibshirani, “A
clinical decision making through implementation of a mo- simple method for estimating interactions between a treat-
lecular tumor board: a German single-center experience,” JCO ment and a large number of covariates,” Journal of the
Precision Oncology, vol. 2, no. 2, pp. 1–16, 2018. American Statistical Association, vol. 109, no. 508, pp. 1517–
[4] T. Ahmadzada, S. Kao, G. Reid, M. Boyer, A. Mahar, and 1532, 2014.
W. Cooper, “An update on predictive biomarkers for treat- [21] J. Fan, H. Lin, and Y. Zhou, “Local partial-likelihood esti-
ment selection in non-small cell lung cancer,” Journal of mation for lifetime data,” The Annals of Statistics, vol. 34,
Clinical Medicine, vol. 7, no. 6, p. 153, 2018. no. 1, pp. 290–325, 2006.
[5] K. V. Ballman, “Biomarker: predictive or prognostic?,” [22] K. Strimbu and J. A. Tavel, “What are biomarkers?,” Current
Journal of Clinical Oncology, vol. 33, no. 33, pp. 3968–3971, Opinion in HIV and AIDS, vol. 5, no. 6, pp. 463–466, 2010.
2015. [23] Personalized Medicine Coalition, “Personalized medicine at
[6] K. P. Pritzker, “Predictive and prognostic cancer biomarkers FDA: 2017 progress report,” January 2019, http://www.
revisited,” Expert Review of Molecular Diagnostics, vol. 15, personalizedmedicinecoalition.org/Userfiles/PMC-Corporate/
no. 8, pp. 971–974, 2015. file/PM_at_FDA_2017_Progress_Report.pdf.
[7] M.-Y. C. Polley, B. Freidlin, E. L. Korn, B. A. Conley, [24] M. K. Wilson, D. Collyar, D. T. Chingos et al., “Outcomes and
J. S. Abrams, and L. M. McShane, “Statistical and practical endpoints in cancer trials: bridging the divide,” The Lancet
considerations for clinical evaluation of predictive bio- Oncology, vol. 16, no. 1, pp. e43–e52, 2015.
markers,” JNCI Journal of the National Cancer Institute, [25] M. Bonetti and R. D. Gelber, “A graphical method to assess
vol. 105, no. 22, pp. 1677–1683, 2013. treatment-covariate interactions using the Cox model on
[8] A. D. Hingorani, D. A. v. d. Windt, R. D. Riley et al., subsets of the data,” Statistics in Medicine, vol. 19, no. 19,
“Prognosis research strategy (progress) 4: stratified medicine pp. 2595–2609, 2000.
research,” BMJ, vol. 346, p. e5793, 2013. [26] D. R. Cox, “Regression models and life-tables,” Journal of the
[9] E. Senkus, S. Kyriakides, S. Ohno et al., “Primary breast Royal Statistical Society: Series B (Methodological), vol. 34,
cancer: ESMO Clinical Practice Guidelines for diagnosis, no. 2, pp. 187–202, 1972.
treatment and follow-up,” Annals of Oncology, vol. 26, no. S5, [27] Y. Liu, W. Jiang, and B. E. Chen, “Testing for treatment-
pp. v8–v30, 2015. biomarker interaction based on local partial-likelihood,”
[10] F. Cardoso, E. Senkus, A. Costa et al., “4th ESO-ESMO in- Statistics in Medicine, vol. 34, no. 27, pp. 3516–3530, 2015.
ternational consensus guidelines for advanced breast cancer [28] R Core Team, R: A Language and Environment for Statistical
(ABC 4),” Annals of Oncology, vol. 29, no. 8, pp. 1634–1657, Computing, R Foundation for Statistical Computing, Vienna,
2018. Austria, 2016.
[11] B. Besse, A. Adjei, P. Baas et al., “2nd ESMO Consensus [29] T. M. Therneau, A Package for Survival Analysis in S, Mayo
Conference on Lung Cancer: non-small-cell lung cancer first- Foundation, Rochester, MN, USA, 2015.
line/second and further lines of treatment in advanced dis- [30] T. M. Therneau and P. M. Grambsch, Modeling Survival Data:
ease,” Annals of Oncology, vol. 25, no. 8, pp. 1475–1484, 2014. Extending the Cox Model, Springer Science & Business Media,
[12] J. J. Chen, T.-P. Lu, Y.-C. Chen, and W.-J. Lin, “Predictive New York, NY, USA, 2000.
biomarkers for treatment selection: statistical considerations,” [31] J. Fox and S. Weisberg, An R Companion to Applied
Biomarkers in Medicine, vol. 9, no. 11, pp. 1121–1135, 2015. Regression, Sage, Thousand Oaks, CA, USA, 2nd edition, 2011.
[32] M. Bonetti, D. Zahrieh, B. F. Cole, and R. D. Gelber, “A small [48] A.-L. Boulesteix, S. Lauer, and M. Eugster, “A plea for neutral
sample study of the STEPP approach to assessing treatment- comparison studies in computational sciences,” PLoS One,
covariate interactions in survival data,” Statistics in Medicine, vol. 8, no. 4, pp. 1–11, 2013.
vol. 28, no. 8, pp. 1255–1268, 2009. [49] D. G. Altman, B. Lausen, W. Sauerbrei, and M. Schumacher,
[33] W.-K. Yip, M. Bonetti, B. F. Cole et al., “Subpopulation “Dangers of using “optimal” cutpoints in the evaluation of
treatment effect pattern plot (STEPP) analysis for contin- prognostic factors,” JNCI Journal of the National Cancer
uous, binary, and count outcomes,” Clinical Trials: Journal of Institute, vol. 86, no. 11, pp. 829–835, 1994.
the Society for Clinical Trials, vol. 13, no. 4, pp. 382–390, [50] M. Bonetti, B. F. Cole, and R. D. Gelber, “Another STEPP in
2016. the right direction,” Journal of Clinical Oncology, vol. 26,
[34] A. A. Lazar, M. Bonetti, B. F. Cole, W.-K. Yip, and no. 22, pp. 3813-3814, 2008.
R. D. Gelber, “Identifying treatment effect heterogeneity in [51] I. Lipkovich, A. Dmitrienko, and B. Ralph, “Tutorial in
clinical trials using subpopulations of events: STEPP,” Clinical biostatistics: data-driven subgroupidentification and analysis
Trials, vol. 13, no. 2, pp. 169–179, 2016. in clinical trials,” Statistics in Medicine, vol. 36, no. 1,
[35] W.-K. Yip, A. Lazar, D. Zahrieh et al., STEPP: Subpopulation pp. 136–196, 2017.
Treatment Effect Pattern Plot (STEPP), R package version [52] P. Royston and W. Sauerbrei, “Interactions between treatment
3.2.0.0, 2018, https://CRAN.R-project.org/package=stepp. and continuous covariates: a step toward individualizing
[36] P. Royston and D. G. Altman, “Regression using fractional therapy,” Journal of Clinical Oncology, vol. 26, no. 9,
polynomials of continuous covariates: parsimonious para- pp. 1397–1399, 2008.
metric modelling,” Applied Statistics, vol. 43, no. 3, pp. 429– [53] T. Ondra, A. Dmitrienko, T. Friede et al., “Methods for
467, 1994. identification and confirmation of targeted subgroups in
[37] P. Royston and W. Sauerbrei, “Interaction of treatment with a clinical trials: a systematic review,” Journal of Bio-
continuous variable: simulation study of significance level for pharmaceutical Statistics, vol. 26, no. 1, pp. 99–119, 2016.
several methods of analysis,” Statistics in Medicine, vol. 32, [54] M. J. Selleck, M. Senthil, and N. R. Wall, “Making meaningful
no. 22, pp. 3788–3803, 2013. clinical use of biomarkers,” Biomarker Insights, vol. 12,
[38] P. Royston and W. Sauerbrei, “Interaction of treatment with a pp. 1–7, 2017.
continuous variable: simulation study of power for several
methods of analysis,” Statistics in Medicine, vol. 33, no. 27,
pp. 4695–4708, 2014.
[39] T. M. Loughin, “A residual bootstrap for regression param-
eters in proportional hazards models,” Journal of Statistical
Computation and Simulation, vol. 52, no. 4, pp. 367–384,
1995.
[40] S. Zhang and B. Chen, lplb: Local Partial Likelihood Bootstrap
(LPLB) Test, R Package Version 0.1, 2016, https://github.com/
statapps/lplb.
[41] J. Cohen, “The cost of dichotomization,” Applied Psychological
Measurement, vol. 7, no. 3, pp. 249–253, 1983.
[42] R. C. MacCallum, S. Zhang, K. J. Preacher, and D. D. Rucker,
“On the practice of dichotomization of quantitative variables,”
Psychological Methods, vol. 7, no. 1, pp. 19–40, 2002.
[43] O. Naggara, J. Raymond, F. Guilbert, D. Roy, A. Weill, and
D. G. Altman, “Analysis by categorizing or dichotomizing
continuous variables is inadvisable: an example from the
natural history of unruptured aneurysms,” American Journal
of Neuroradiology, vol. 32, no. 3, pp. 437–440, 2011.
[44] P. Royston and W. Sauerbrei, “Two techniques for in-
vestigating interactions between treatment and continuous
covariates in clinical trials,” The Stata Journal: Promoting
communications on statistics and Stata, vol. 9, no. 2,
pp. 230–251, 2009.
[45] P. Ringleb, J. Allenberg, J. Berger et al., “30 day results from
the SPACE trial of stent-protected angioplasty versus carotid
endarterectomy in symptomatic patients: a randomised non-
inferiority trial,” The Lancet, vol. 368, no. 9543, pp. 1239–1247,
2006.
[46] H.-H. Eckstein, P. Ringleb, J.-R. Allenberg et al., “Results of
the stent-protected angioplasty versus carotid endarterectomy
(SPACE) study to treat symptomatic stenoses at 2 years: a
multinational, prospective, randomised trial,” The Lancet
Neurology, vol. 7, no. 10, pp. 893–902, 2008.
[47] B. Haller, H.-H. Eckstein, P. A. Ringleb, and K. Ulm, “In-
vestigation of age–treatment interaction in the space trial
using different statistical approaches,” Journal of Applied
Statistics, vol. 46, no. 9, pp. 1689–1701, 2019.
Hindawi
https://doi.org/10.1155/2019/7370231
Research Article
A Data-Driven Hybrid Three-Stage Framework for Hospital Bed
Allocation: A Case Study in a Large Tertiary Hospital in China
Li Luo ,1 Jialing Li ,1 Xueru Xu ,1 Wenwu Shen ,2 and Lin Xiao 1
1
Business School of Sichuan University, No. 24 South Section 1, Yihuan Road, Chengdu, China
2
West China Hospital of Sichuan University, No. 17 People’s South Road, Chengdu, China
Correspondence should be addressed to Xueru Xu; 18200291987@163.com and Wenwu Shen; wenwu_shen@163.com
Received 24 December 2018; Revised 6 March 2019; Accepted 2 April 2019; Published 2 May 2019
Guest Editor: Giedrius Vanagas
Copyright © 2019 Li Luo et al. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Beds are key, scarce medical resources in hospitals. The bed occupancy rate (BOR) amongst different departments within large
tertiary hospitals is very imbalanced, a situation which has led to problems between the supply of and the demand for bed
resources. This study aims to balance the utilization of existing beds in a large tertiary hospital in China. We developed a data-
driven hybrid three-stage framework incorporating data analysis, simulation, and mixed integer programming to minimize the
gaps in BOR among different departments. The first stage is to calculate the length of stay (LOS) and BOR of each department and
identify the departments that need to be allocated beds. In the second stage, we used a fitted arrival distribution and median LOS as
the input to a generic simulation model. In the third stage, we built a mixed integer programming model using the results obtained
in the first two stages to generate the optimal bed allocation strategy for different departments. The value of the objective function,
Z, represents the severity of the imbalance in BOR. Our case study demonstrated the effectiveness of the proposed data-driven
hybrid three-stage framework. The results show that Z decreases from 0.7344 to 0.0409 after re-allocation, which means that the
internal imbalance has eased. Our framework provides hospital bed policy makers with a feasible solution for bed allocation.
1. Introduction Amongst all of the medical institutions, China’s large ter-

tiary hospitals, classed as Class III according to the classi-
The inherent difference between limited resources for fication standards (Appendix), are facing the most serious
healthcare and steadily increasing demands occurs all over imbalance between admissions and bed resources (Table 1).
the world and is particularly serious in developing countries. The number of hospital beds and individuals hospitalized in
According to a research report from the World Health Class III hospitals increased to 2,213,718 and 76,860,000,
Organization (WHO) and World Bank Groups, at least 400 respectively. Bed occupancy rates (BORs) reached 98.8% in
million people worldwide cannot receive one or more basic 2016.
health services [1]. We found that the imbalance between supply and de-
This differential is particularly apparent with respect to mand in large tertiary (Class III) hospitals is greater than
bed resources. Although hospital bed numbers have in- that in the other two classes of hospitals. Hospital managers
creased greatly in recent years, this increase cannot cope urgently need to find solutions to alleviate bed shortages.
with the growth rate of admission demand in China. Hospital administrators typically address this issue in two
According to the report “Statistical Communique on the ways: by improving utilization of existing beds or by
Development of China’s Health and Family Planning Pro- expanding capacity. The first involves the complex task of
gram 2016” [2], the number of hospitalizations across the strategically allocating the proper amount of beds for each
nation’s medical and health institutions was 227.28 million set of care types. A hospital that is unable to find an optimal
and the annual hospitalization rate was 16.5%. There were allocation may acquire additional beds. However, an ex-
7.410 million beds in medical institutions across the country. pansion, which is desirable for one hospital, may not be
Table 1: Hospital beds, hospitalization, and BOR of China’s large tertiary hospitals in 2015 and 2016.
Hospital beds Hospitalization BOR
Hospital level
2015 2016 2015 2016 2015 (%) 2016 (%)
Class III 2,047,819 2,213,718 68,290,000 76,860,000 98.80 98.80
Class II 2,196,748 2,302,887 71,210,000 75,700,000 84.10 84.20
Class I 481,876 517,837 9,650,000 10,390,000 58.80 58
advantageous from the perspective of the public planner [3] research on the allocation of beds among different de-
and may even have some negative consequences, such as partments or wards.
doctor and nurse work overloads, decreases in medical care A range of operational research (OR) methods have been
quality, and aggravation of medical conflicts [4]. Therefore, developed and applied to problems of healthcare resource
from the perspective of the sustainability of medical reallocation [15], especially bed resources. These methods
sources, the best option is to improve the utilization of bed include queuing theory [16, 17], simulation [18, 19], goal
resources. programming [20, 21], and mathematical programming
The BOR of the Class III hospitals is usually as high as [22–24]. Most of the current research uses a single method
98%. However, the utilization rate of beds in different de- for each study. However, the premise of queuing theory is
partments of a Class III hospital can be very different. West based on very strong assumptions, and it is difficult to apply.
China Hospital (WCH) is typical of such hospitals (see Mathematical programming can flexibly add constraints and
Section 3.1 for more details). The availability of beds for change the objective function according to specific condi-
patient hospitalization services is excessive in some units and tions, and it is more popular. Therefore, a hybrid of different
scarce in others. This discrepancy between bed availability in methods is more conducive to solving the problem. Simu-
different departments leads to the overcrowding of some lation and mathematical programming have become in-
departments and the idleness of other departments. This creasingly popular approaches to allocate resources in health
internal imbalance has further worsened the shortage of care. Tontarski [25] utilized simulation-based optimization
hospital resources. and mathematical programming for solving complex nurse-
A possible solution to this problem is to allocate the scheduling problems. Studies on the bed allocation problem
number of beds among different departments in such a way using mathematical programming, especially combined with
as to increase the utilization rate of beds and alleviate the simulation, are relatively few.
shortage of bed resources as much as possible. There has In short, research into bed allocation mostly focuses on
been considerable research into the allocation of bed re- the study of the number of beds in a single department. The
sources. The study of bed resource allocation can be divided problem of how to allocate beds in different departments has
into two main approaches. not been fully studied. Furthermore, most of the current
One approach is to assign the optimal number of beds in research on bed allocation is based on a single method such
a single department or care units, such as surgery units [5], as queuing theory [26] or simulation [27]. There are few
intensive care units [6, 7], and obstetrics departments [8]. hybrid models which integrate data analysis, simulation, and
For instance, Akkerman and Knip [5] planned the optimal mathematical programming. We propose a solution to the
number of beds for cardiac surgery with the goal of reducing bed allocation problem at an operational level. We develop a
patient waiting time. Oerlemans et al. [9] sent an online data-driven hybrid three-stage framework incorporating
questionnaire to all ICU physician members in 90 hospitals data analysis, simulation, and mixed integer programming
of the Dutch Society for Intensive Care, the results of which (MIP) to determine the optimal bed allocation strategy. The
can be used to improve decision-making regarding alloca- first stage is to select departments for allocation according to
tion of ICU resources. Devapriya et al. [10] proposed the the relationship between the number of beds and the BOR
strategic bed analysis model, which is a discrete-event and have this result approved by hospital management. In
simulation model created after a thorough analysis of pa- the second stage, we used a simulation model to calculate the
tient flow and data from Geisinger Health Systems (GHS). BOR for different numbers of hospital beds. We thus derive
Ridge et al. [11] investigated the problem of hospital bed the functional relationship between the two variables. The
planning in the intensive care unit. Romanin-Jacur and third stage is to find the best number of beds in five de-
Facchin [12] studied the ward planning of the intensive partments using a MIP model. Our study aims to alleviate
surgical department and the pediatric semiintensive care
shortage of beds by balancing the utilization of existing beds
unit.
without increasing the number of beds in a large hospital in
Another approach to the problem focuses on the number
China. Overall, the contributions of this study are as follows:
of hospital beds throughout the whole hospital. For example,
Akcali et al. [13] planned the best use of hospital beds in the (1) Our framework is data driven, making the allocation
entire hospital with the goal of minimizing the total cost. strategy more rational. (i) Using real data from the
Utley et al. [14] determined the reasonable number of beds hospital, data analysis is used to determine the op-
for elective patients in the whole hospital in the case of a very timal department needs to allocate beds; (ii) the
low rate of patient cancellation. However, there is little simulation model is used to simulate the
corresponding BOR under different bed allocation 1 to December 31, 2013. It includes the time of each patient
scenarios. The relationship between the hospital beds admission and discharge, demographic information, and
and the BOR is derived from the data, thereby department information and has a total of 243,685 admis-
providing data-driven personalized constraints for sion registrations and 167,843 discharge records.
each department.
(2) The objective function of the MIP model is tailored 2.3. Methods. The aim of this research is to balance the BOR
to Chinese needs. Because the beds in large hospitals of each department by allocating the hospital beds to de-
in China have been overloaded, we are not blindly partments, using a fixed number of existing beds, and
reducing the BOR of beds but are keeping the keeping the bed utilization rate of each department at a
number of beds in a reasonable range. reasonable level. Hence, we proposed a data-driven hybrid
(3) This paper provides a general framework for the three-stage framework to solve this problem. The overall
allocation of beds in the Chinese context. Different approach is shown in Figure 2.
hospitals can modify the objective function of the
(1) Stage I (data preliminaries): we selected the key
model and appropriately add or reduce constraints
departments by analyzing their current BOR and
according to their needs. Our model provides a
number of beds.
reference for hospital management to effectively
manage hospital bed resources. (2) Stage II (construction of constraint conditions): Simio
software [28] was used to establish a simulation model
The rest of the paper is organized as follows. In Section 2, to obtain the different scenarios of the beds and cor-
we briefly introduced the background of the case hospital responding BOR. We then determined the relationship
(WCH) and the process of the data-driven hybrid three- between the BOR and the number of beds through data
stage framework. Taking WCH as a case study, we applied fitting, which is one of the constraints of the Stage III.
the framework proposed in Section 2 to WCH, and these
(3) Stage III (construction of model): we established a
analyses of the framework are shown in Section 3. In Section
MIP model to minimize the gap in BOR among
4, we discuss the results of the paper. Finally, Section 5
different departments. We applied the genetic al-
concludes the paper and indicates some directions for future
gorithm (GA) to solve this model, since GA is one of
research.
the best tools for satisfactory solution with advan-
tages like good convergence, low computational
2. Materials and Methods complexity, high robustness, and so forth [29].
2.1. Study Hospital. From a macro perspective, the overall
BOR in China’s tertiary hospitals is very high. However, at 2.3.1. Stage I: Data Preliminaries
an individual level, the availability of hospital beds in dif-
ferent departments in individual hospitals is uneven. This (1) Calculation of Length of Hospital Stay. Length of hospital
phenomenon exists in almost all of the tertiary hospitals in stay (LOS) indicates the number of days the patient spent in
China. It is particularly serious in WCH, a tertiary hospital a hospital bed. We made the assumption that the LOS can be
which is located in Chengdu, Sichuan province. In order to considered as a constant [30]. We have got 243,685 ad-
rationally manage beds, the Admission Service Center mission records and 167,843 discharge records form ASC.
(ASC), a bed planning organization, was established in 2011. The LOS is calculated as the discharge date for each patient
It manages 2956 beds and 28 specialized care departments. minus the date of admission registration in the ASC. The
After our survey and data analysis (Figure 1), we found that sum of the days of all hospitalized patients is an important
the allocation of beds among the 28 departments is not parameter for calculating the BOR. This paper uses the data
balanced. For example, the BOR of W3 is as high as 122%, from 2013 1/1 to 2013/12/31. We divided the patients into
while others, such as W12, are only 64.70%. The reason why three types by the discharge date as follows and calculated
the BOR is over 100% is that the extra beds are involved in the LOS in 2013 for each group.
the calculation process. When the number of inpatients (i) Type I: patients discharged during 2013. Those
exceeds the number of fixed beds, the hospital will add patients registered in the ASC during or before 2013.
additional beds to meet the demand, and these beds are often Their actual LOS during the 2013 is called LOS1. For
arranged in the corridor. This imbalance further leads to example, the actual LOS of the patient who regis-
inefficiency and waste of hospital bed space, which in turn tered before 2013 is discharge date minus January 1,
exacerbates the shortage of hospital beds. 2013.
We focus on balancing the BOR of departments by
(ii) Type II: those patients who left the hospital in 2014,
redistributing beds to improve the utilization rate of resources.
who registered before 2013 or during 2013. Their
This study takes WCH as an example to provide a feasible
LOS equals to December 1, 2013, minus registered
solution for the shortage of hospital beds in large hospitals.
date or December 31, 2013, minus January 1, 2013,
which is named LOS2.
2.2. Data Collection. This study used data from the Hospital (iii) Type III: because some discharge data are missing,
Information System of the ASC for the period from January there are patients who are recorded as having been
250 250.00%
200 200.00%
150 150.00%
100 100.00%
50 50.00%
0 0.00%
w1
w2
w3
w4
w5
w6
w7
w8
w9
w10
w11
w12
w13
w14
w15
w16
w17
w18
w19
w20
w21
w22
w23
w24
w25
w26
w27
w28
Hospital bed
BOR
Figure 1: Hospital beds and BOR of 28 departments in West China Hospital.
Data LOS Department

Data source preprocessing BOR selecting Stage I
Five departments
Bed Arrival rate

Data fitting II Simulation Data fitting I Stage II
BOR LOS distribution
Quadratic curve function

of bed and BOR
Stage III
Mixed integer programming model
Figure 2: Methodology: process of the data-driven hybrid three-stage framework.
discharged from hospital in 2013 but had no ad- constraint condition in building the mixed integer pro-
mission records. These patients cannot be ignored. gramming model in Stage III. It proceeds in three steps:
Hence, the number of Type III patients is the total
(1) We used the EASY-FIT [31], a professional data
number of discharged patients in one department
fitting software, to fit the patient’s arrival distribution
in 2013 minus the number of Type I. We cannot
and LOS distribution of each department
directly calculate the LOS of Type III patients, so in
this study, we used the median of the LOS of Type I (2) The fitted arrival distribution and LOS distribution
patients as the value for the LOS of Type III pa- were used as the input to the simulation model,
tients, which we called LOS3. which identified the relationship between the
number of beds and the BOR
(3) We fitted the relationship between the number of
(2) The Calculation of BOR. We calculated each department’s
beds and the BOR via IBM SPSS Statistics V21 and
BOR according to the following formulas:
obtained their functional relationship
all patients’ LOS � LOS1 + LOS2 + LOS3, (1)
all patients’ LOS 2.3.3. Stage III: Construction of the Model. In order to
BOR � . (2)
LOS that all beds can provide thoroughly understand the hospital bed allocation problem,
it is necessary to describe the characteristics of the problem
in order to implement them in an appropriate mathematical
2.3.2. Stage II: Construction of Constraint Conditions. model. We define parameters and variables of the model:
The Simio software was used to build simulation models (1) Parameters
using data from different departments. Changes in the
Kij : the ward type j for department i. There are
number of beds, the number of hospitalized patients, and the
three inpatient ward types.
LOS of patients with different bed numbers were simulated
and used to calculate different BOR indexes. The relationship Ki1 : the number of single-bed wards in department i.
between hospital bed numbers and BOR was an important Ki2 : the number of double-bed wards in department i.
Ki3 : the number of three-bed wards in department i constant. Function (6) ensures that there will be an upper
limit and a lower limit for the beds in each department.
Li : lower bound of number of beds in department i.
Functions (7) and (8) impose restrictions on the ward type
Ui : upper bound of number of beds in department i. and patient gender. Each ward has either one, two, or three
Ct : the total number of beds of all the departments. beds. The distinction between male and female wards, and
BORi (Ci ): the BOR of department i, when the the number of ward types in each department should not be
number of beds of department i is Ci . The re- less than two. For example, a single-bed ward has at least two
lationship functions between BORi (Ci ) and Ci of wards so that a male patient can live in one room and a
department i can be obtained from the results of female patient can live in another single room. The other two
Stage II. Here, we assume that the two variables are ward types have the same conditions. Both male and female
quadratic functions. patients can decide which type of ward to live in, ensuring
the fair treatment of patients.
(2) Variable
Ci : the number of beds of department i and 3. Results
Ci � Ki1 + 2Ki2 + 3Ki3 ; Ci is a positive integer.
There are n departments in WCH, but i de- 3.1. Stage I: Data Preliminaries. The LOS of 28 departments
partments (i ∈ 1, 2, . . . , n) need to be allocated beds. can be calculated by function (1), and the BOR of the 28
Our main decision variable is Ci , which represents departments can be calculated by function (2). Table 2 shows
the number of beds for department i. Our goal was to the number of beds and BOR in 28 departments of the
balance the BOR of each department; hence, our current WCH. The BOR varies from 60.7% to 195% among
objective function is to minimize the total gap be- the 28 departments. Some literature indicates [32, 33] that
tween BOR of each department and their average the optimal range for BOR is between 85% and 90%. Based
BOR. We developed a MIP model as follows: on this estimation, we divided these departments into three
groups:
(3) Objective function
􏼌 􏼌􏼌 (1) Group A: BOR is less than 85%. For example, W9
n 􏼌􏼌􏼌 n 􏼌􏼌
􏼌􏼌 BOR C 􏼁 owns 236 beds, but its BOR is only 75.2%.
min(Z) � 􏽘􏼌􏼌BORi Ci 􏼁 −⎛􏽘 ⎝ i i ⎞􏼌􏼌􏼌.
⎠ (3)
􏼌 n 􏼌􏼌 (2) Group B: BOR is greater than 90%. For example, the
i�1 􏼌 i�1 􏼌
BOR of W6 reaches as high as 102%, but it only has
72 beds.
(4) Constraints
(3) Group C: BOR is between 85% and 90%. Their bed
numbers and BOR are within the normal range,
BORi Ci 􏼁 � βi + ai Ci + bi Ci 2 , i � 1, 2, . . . n, (4) compared to groups A and B.
It is clear that there are serious imbalances in BOR
n
􏽘 Ci � Ct , i � 1, 2, . . . n, between departments. To solve this problem, we interviewed
(5)
i�1
a hospital manager, three other managerial assistants, and
medical physicians of the ASC. We choose five departments
Li ≤ Ci ≤ Ui , i � 1, 2, . . . , n, (6) (W9, W10, W19, W6, and W27) from groups A and B to
solve the problem of bed allocation by applying the
framework mentioned in Section 2.3.
Ci � Ki1 + 2Ki2 + 3Ki3 , i � 1, 2, . . . , n, (7)
Kij ≥ 2, i � 1, 2, . . . , n, j � 1, 2, 3. (8) 3.2. Stage II: Construction of Constraint Conditions. After

perprocessing the data and selecting the departments, we
Because of the imbalance between different specialty care fitted the distribution of the patient arrival rate and LOS of
departments, we aimed to balance the BOR of various de- the five departments using EASY-FIT, and the results are
partments without adding extra beds. In objective function shown in Table 3.
(3), (􏽐ni�1 BORi (Ci ))/n is the average BOR of n departments We obtained the distribution of arrival rates for all five
and Z is the sum of the gap between the BOR of the n departments. The fitting of LOS is not ideal. Five de-
departments and the average of their BOR. The purpose of partments do not display any distribution. We took the
function (3) is to minimize the total gap in BOR between median of the LOS as their distribution. We used W9 as the
different departments and average BOR. example from which we can build the simulation model.
The constraint described by function (4) means the Figure 3 describes the simulation model of W9 in Simio.
functional relationship between BOR and the number of We set up a patient entity, called Patients, and a patient
beds, which is calculated in Stage II. Here, we assume source called Source1, in the Simio software. We let Source1
that the two are quadratic functions: BORi (Ci ) � βi + associate with Patients and set the arrival rate to obey the
ai Ci + bi Ci 2 , where βi is a constant and ai and bi are co- Johnson distribution (0.025, 0.803, −8.16, 85.98).
efficients (for more details, see Section 3.2); function (5) We built a Server1 to represent beds. Its Service capacity
means that the total number of beds of the n departments is was set to the current number of beds (236), and service time
Table 2: Department information summary.

Department W7 W8 W9 W10 W12 W15 W16 W17 W19
Beds 72 84 236 168 48 60 86 156 54
BOR (%) 83.4 75.7 75.2 71.8 64.7 70 60.7 71.1 70
A
Department W24 W23 W25
Beds 162 153 114
BOR (%) 75.60 84.40 67.60
Department W1 W3 W4 W5 W6 W11 W13 W18 W21
Beds 72 108 84 108 72 72 170 172 72
BOR (%) 94.0 122 110 91 102 95.8 93.5 100 167
B
Department W26 W27 W26
Beds 91 114 91
BOR (%) 195 137 195
Department W1 W2 W20 W28
C Beds 84 66 72 114
BOR (%) 89.6 87.7 90 88.1
Table 3: Results of arrival rate distribution and LOS.

Departments Department type Arrival rate LOS
W6 B Johnson SB (0.289, 0.983, −3.34, 35.11) Median � 9
W9 A Johnson SB (0.025, 0.803, −8.16, 85.98) Median � 10
W10 A Uniform (−0.72, 43.75) Median � 12
W19 A Uniform (−3.2, 92.36) Median � 10
W27 B Uniform (−3.2, 92.36) Median � 5
Figure 3: Simulation model of the W9.
was set to 10 days. We calculated the proportion of hospi- for the Server1 and calculated the BOR for different
talizations (43%) based on the number of hospital admis- scenarios.
sions (12712) and discharges (5478). Because W9 belongs to group A, we needed to reduce the
There are two leaving routes in the simulation model, number of beds and increase the BOR. We should therefore
namely, Sink1 and Sink2. Sink1 represents the event that a reduce the number of beds in Server1. In Table 4, we list the
patient leaves the hospital after being served by Server1. The number of beds and the corresponding BOR situation for
weight from Source1 to Server1 is 43%. Sink2 indicates that a W9. The results of the other four departments are presented
patient who was not admitted to the hospital left the hospital in Tables 5–8.
directly through Sink2; the weight from Source1 to Sink2 is We used the number of beds as an independent variable
57%. and the BOR as the dependent variable based on the results
In order to validate the model, we run the model for a in Table 4. The graph of BOR changing with the number of
simulated year. The result was that the total number of beds is shown in Figure 4. It is difficult to intuitively obtain
discharges was 5984, and the BOR was 69.5%. Compared the relationship equation between the two from the graph, so
with real data, the difference is 9.2% and 7.5%, respectively. we selected eight kinds of curve functions—linear, loga-
The error was acceptable. Then, we set different parameters rithm, quadratic, composite, power, growth, exponential,
Table 4: Beds and corresponding BOR of W9 in the simulation Table 7: Beds and corresponding BOR of W27.
model.
Beds Discharge BOR (%)
Beds Discharge BOR (%) 114 8196 98.50
92 3266 97.3 130 9332 98.30
110 3896 97 145 10345 97.70
123 4351 96.9 149 10482 96.37
135 4771 96.8 152 10527 94.90
148 5226 96.7 160 10640 91.10
160 5646 96.68 168 10674 87
172 5926 94.4 175 10681 83.60
180 5982 91.1 182 10688 80.40
185 5984 88.6 190 10696 77.10
198 5984 82.8 205 10711 71.60
210 5984 78.1 220 10726 66.80
223 5984 73.5
236∗ 5984 69.5
250 5984 65.6
∗
Table 8: Beds and corresponding BOR of W19.
The current number of hospital beds of W9.
Beds Discharge BOR (%)
30 995 97.2
Table 5: Beds and corresponding BOR of W10. 35 1153 97
Beds Discharge BOR (%) 36 1189 96.8
39 1225 93
68 2011 97.2
38 1165 90.3
78 2301 97
40 1258 92.2
88 2585 96.6
45 1246 80.9
98 2865 96.1
47 1230 77.4
108 3145 95.7
50 1246 73.05
113 3201 93.1
52 1252 70.6
118 3216 89.6
54 1274 68.9
128 3246 83.4
138 3276 78
148 3299 73.3
158 3306 68.8 BOR
100.00%
168 3306 64.7 90.00%
178 3306 61.1 80.00%
70.00%
60.00%
50.00%
Table 6: Beds and corresponding BOR of W6. 40.00%
30.00%
20.00%
Beds Discharge BOR (%) 10.00%
57 2279 98.6 0.00%
80 100 120 140 160 180 200 220 240 260
62 2474 98.4
72 2864 98.1 Figure 4: BOR curve for different numbers of beds. The abscissa is
82 3254 97.8 the number of beds, and the ordinate is BOR.
95 3756 97.4
100 3931 96.9
105 4106 96.4 relationship between BOR1 (C1 ) and the bed C1 from
110 4281 96
Table 9; the equation is: BOR1 (C1 ) � 0.721 + 0.004C1 −
115 4438 95.1
120 4516 92.8
0.0000195C21 .
125 4538 89.5 Similar to the analysis process of W9, we obtained the
130 4538 86.1 quadratic functional relationships between the BOR and the
140 4538 80 bed number of the other four departments (Tables 10–13).
Hence, the relationship function between beds and BOR of
department i is expressed as follows:
and logistic—with which we can attempt to fit their func-
tional relationships. We used the value of R2 to determine BORi Ci 􏼁 � βi + ai Ci + bi Ci2 , (9)
which relationship function between bed and BOR of W9
had the best fit (Table 9). Since the R2 value of the quadratic where BORi (Ci ) represents the bed occupancy rate of de-
function was the best, at 0.977, we decided that the quadratic partment i, Ci is the number of beds in department i, βi is a
function best describes the relationship between the bed constant, and ai and bi are coefficients. We have obtained
numbers and the BOR in W9. We can derive the functional five equations, respectively, for five departments. They are
Table 9: Relationship between bed numbers and BOR in W9.

Model Parameters
Function
R2 F df1 df2 Sig. Constant b1 b2 b3
Linear 0.836 61.326 1 12 0.000 1.253 −0.002
Logarithm 0.724 31.538 1 12 0.000 2.550 −0.328
Quadratic 0.977 230.860 2 11 0.000 0.721 0.004 0.0000195
Composite 0.820 54.831 1 12 0.000 1.367 0.997
Power 0.704 28.548 1 12 0.000 6.455 −0.392
Growth 0.820 54.831 1 12 0.000 0.313 −0.003
Exponential 0.820 54.831 1 12 0.000 1.367 −0.003
Logistic 0.820 54.831 1 12 0.000 0.732 1.003
Table 10: Relationship between bed numbers and BOR in W10.

Model Parameters
Function
Linear 0.919 125.130 1 11 0.000 1.298 −0.004
Logarithm 0.837 56.598 1 11 0.000 2.813 −0.413
Quadratic 0.973 181.654 2 10 0.000 0.895 0.003 −2.854E − 5
Composite 0.907 107.562 1 11 0.000 1.468 0.995
Power 0.817 49.021 1 11 0.000 9.578 −0.513
Growth 0.907 107.562 1 11 0.000 0.384 −0.005
Exponential 0.907 107.562 1 11 0.000 1.468 −0.005
Logistic 0.907 107.562 1 11 0.000 0.681 1.005
Note. The higher the R2, the better the function model. In the W10 model fitting, the fitting degree of the quadratic curve is the best, and the quadratic curve is
directly selected. The bed number of W10 is C2, and the bed utilization rate is BOR2 (C2). According to the estimated value of the parameter,
BOR2 (C2 ) � 0.895 + 0.003C2 − 0.00002854C22 .
Table 11: Relationship between beds and BOR in W6.

Model Parameters
Function 2
R F df1 df2 Sig. Constant b1 b2 b3
Linear 0.664 21.741 1 11 0.001 1.118 −0.002
Logarithm 0.558 13.883 1 11 0.003 1.616 −0.147
Quadratic 0.956 107.465 2 10 0.000 0.687 0.008 −4.932E − 5
Composite 0.645 19.978 1 11 0.001 1.140 0.998
Power 0.539 12.849 1 11 0.004 1.966 −0.161
Growth 0.645 19.978 1 11 0.001 0.131 −0.002
Exponential 0.645 19.978 1 11 0.001 1.140 −0.002
Logistic 0.645 19.978 1 11 0.001 0.877 1.002
Note. The higher the R2, the better the function model. When the W6 model is fitted, it has the same outcome as W9. The quadratic curve is also selected. The
bed number of W6 is C3, and the bed rate is BOR3 (C3). According to the estimated value of the parameter, BOR3 (C3 ) � 0.687 + 0.008C3 − 0.000049323C23 .

Model Parameters
Function
Linear 0.932 137.447 1 10 0.000 1.446 −0.003
Logarithm 0.882 74.410 1 10 0.000 3.673 −0.550
Quadratic 0.967 131.496 2 9 0.000 0.901 0.003 −2.007E − 5
Composite 0.964 119.131 2 9 0.000 1.094 0.000 −2.585E − 6 −3.039E − 8
Power 0.924 121.995 1 10 0.000 1.720 0.996
Growth 0.866 64.725 1 10 0.000 24.281 −0.655
Exponential 0.924 121.995 1 10 0.000 0.542 −0.004
Logistic 0.924 121.995 1 10 0.000 1.720 −0.004
Note. The higher the R2, the better the function model. In the W27 model fitting, the fitting of the quadratic curve is the best, and the quadratic curve is
selected. The bed number of W27 is C4, and the bed utilization rate is BOR4 (C4). According to the estimated value of the parameter,
BOR4 (C4 ) � 0.901 + 0.003C4 − 0.000020066C24 .

Model Parameters
Function
Linear 0.947 159.652 1 9 0.000 1.450 −0.014
Logarithm 0.915 96.619 1 9 0.000 3.011 −0.579
Quadratic 0.966 114.721 2 8 0.000 0.911 0.012 −3.054E − 5
Composite 0.964 108.277 2 8 0.000 1.095 0.000 −5.576E − 5 −1.687E − 6
Power 0.944 151.827 1 9 0.000 1.736 0.983
Growth 0.908 88.465 1 9 0.000 11.291 −0.695
Exponential 0.944 151.827 1 9 0.000 0.552 −0.017
Logistic 0.944 151.827 1 9 0.000 1.736 −0.017
Note. The higher the R2, the better the function model. In the W19 model fitting, the fitting of the quadratic curve is the best, and the quadratic curve is
selected. The bed number of W19 is C5, and the bed utilization rate is BOR5 (C5). According to the estimated value of the parameter,
BOR5 (C5 ) � 0.911 + 0.012C5 − 0.0003054C25 .
used as constraints for the mixed integer programming 114 ≤ C4 ≤ 182, (20)
model in Stage III.
30 ≤ C5 ≤ 182, (21)
3.3. Stage III: A Mixed Integer Programming Model. After
data analysis and simulation in the first two stages, we Ci � Ki1 + 2Ki2 + 3Ki3 , i � 1, 2, 3, 4, 5, (22)
obtained the parameter values in equations (3)–(8), in-
cluding the five selected departments and established the Kij ≥ 2, i � 1, 2, 3, 4, 5, j � 1, 2, 3. (23)
quadratic function relationship between the bed numbers Constraints (11)–(15) are quadratic functions of the
and the BOR. We then applied these parameters to equations number of beds and the BOR of the five departments. After
(3)–(8) to solve the model. The specific MIP model is as we obtain the number of beds in each department, we can
follows: calculate their BOR by formulas (11)–(15). Constraint (16)
Objective function states the total number of beds in five departments. Con-
straints (17)–(21) limit the upper bound and lower bound on
bed numbers of each department. Constraints (22) and (23)
􏼌􏼌 􏼌􏼌 restrict the ward type and patient gender. There are three
􏼌􏼌
5 5
BORi Ci 􏼁⎠ 􏼌􏼌
􏼌􏼌 ⎝
⎛ ⎞ 􏼌􏼌 ward types for each department in WCH. So, the number of
min(Z) � 􏽘􏼌􏼌BORi Ci 􏼁 − 􏽘 􏼌􏼌. (10)
i�1 􏼌
􏼌 i�1
5 􏼌􏼌 beds in department i is the sum of the total number of beds
from those three types. In order to distinguish the male and
female wards, the number of each department type must be a
Constraints positive integer and should not be less than two. We used the
genetic algorithm [34] to solve the MIP model. The genetic
BOR1 C1 􏼁 � 0.721 + 0.004C1 − 0.0000195C21 , (11)
algorithm is run on a personal computer with an Intel
®
™
Core i7-7700 CPU, a 3.60 GHz z Intel processor, and
8.0 GB RAM. The elapsed time is 77.652611 seconds.
BOR2 C2 􏼁 � 0.895 + 0.003C2 − 0.00002854C22 , (12) We analyzed the results from three aspects:
(1) Initial bed allocations and optimal bed allocations
BOR3 C3 􏼁 � 0.687 + 0.008C3 − 0.000049323C23 , (13)
based on our model: as shown in Figure 5, the blue
histograms represent the Initial bed allocation,
BOR4 C4 􏼁 � 0.901 + 0.003C4 − 0.000020066C24 , (14) which is the current hospital bed number. The op-
timal bed allocations from our model are represented
BOR5 C5 􏼁 � 0.911 + 0.012C5 − 0.0003054C25 , (15) by the yellow histograms. Figure 5 shows the optimal
bed allocation strategy: the bed of the W9 reduces
5 from 236 to 166, W10 from 168 to 121, and W19
􏽘 Ci � 644, (16) from 54 to 44; W6 increases from 72 to 135 and W27
i�1 from 114 to 178.
(2) Initial BOR and optimal BOR: Figure 6 shows the
92 ≤ C1 ≤ 210, (17)
corresponding BOR after optimization. The W9
increases from 69.5% to 84.76%, W10 from 64.7% to
68 ≤ C2 ≤ 138, (18) 84.01%, W19 from 68.9% to 84.78%; on the contrary,
W6 decreases from 98.1% to 86.81% and W27 from
57 ≤ C3 ≤ 140, (19) 98.5% to 83.48%. Blue lines represent the original
250 236 Table 14: Optimal number of beds and different combinations of
departments.
200 178
166 168 Optimal Different feasible combinations of
Hospital beds
Department
150 135 number of beds Kij : (Ki1 , Ki2 , Ki3 )
121 114
100
(K11 , K12 , K13 ) : (26, 40, 20), (22,
72 W9 166
54
42, 20), (36, 20, 30), . . .
44 (K21 , K22 , K23 ) : (21, 20, 20), (21,
50
W10 121
35, 10), (31, 30, 10), . . .
0 (K31 , K32 , K33 ) : (35, 20, 20), (20,
W9 W10 W6 W27 W19 W6 135
40, 7), (25, 10, 30), . . .
Initial bed allocation (K41 , K42 , K43 ) : (78, 20, 20), (10,
W27 178
Optimal bed allocation 39, 30), (40, 39, 20), . . .
(K51 , K52 , K53 ) : (14, 6, 6), (4, 5,
Figure 5: Initial bed allocation strategy and optimal bed allocation W19 44
10), (19, 5, 5), . . .
strategy.
120.00%
98.10% 98.50% utilization of bed resources in different departments. The
100.00% 84.76% 84.01% 86.81% 83.48% 84.78% availability of beds for hospital care is excessive in some cases
80.00% 69.50%
64.70% 68.90% and scarce in others. This phenomenon has caused many
problems for hospitals. For example, some hospital wards
BOR
60.00%
are always overcrowded, while others are underloaded. Some
40.00% scheduled patient admissions are delayed or even transferred
20.00% to other hospitals, and some patients are hospitalized in
0.00%
inappropriate wards which are unsuited to their pathologies,
W9 W10 W6 W27 W19 with the risk of a lower quality care and a greater chance of
Initial BOR infection [35].
Optimal BOR To relieve this imbalance, we propose a data-driven
hybrid three-stage framework combining multiple
Figure 6: Results of initial BOR and optimal BOR.
methods to produce a feasible bed allocation strategy since it
BOR, and yellow lines represent the optimal BOR. We is difficult to allocate beds among all 28 departments in the
found that the maximum and minimum BORs are whole WCH. We selected five departments (W9, W10, W6,
98.1% and 64.7%. The maximum difference of BOR is W27, and W19) through data analysis and survey interview.
33.8% but change to 2.80% after optimization. W9, W10, and W19 are departments that have many beds
with a low BOR while W6 and W27 have few beds with high
(3) Objective function value: our objective value, Z,
BOR. For Stage II, we developed a generic discrete-event
represents the imbalance degree of bed utilization
simulation model. We fitted the relationship function be-
between various departments. For baseline bed al-
tween BOR and beds of each wards via the simulation model.
location, the initial value of the objective function Z
In Stage III, we developed a MIP model to minimize the
is as high as 0.7344. After optimization, the optimal
imbalance in BOR. The results of Stage II are incorporated
value of Z is 0.0409, indicating that our optimization
into the MIP model as one of the key constraints. We also
reduced the severity of the imbalance.
considered other constraints, such as ward types (single,
Finally, we can get a combination of beds in different double, and three-bed wards) and upper and lower bounds
wards based on the number of optimal beds (Table 14). In on the number of beds.
formula (22), Ci depends on the value of Kij , that is to say, Our data-driven hybrid three-stage framework produces
the combinations of Kij produce different Ci values. For a flexible allocation strategy for hospital bed management.
example, the optimal number of beds for W9 determined by Our research helps to improve the utilization of medical
our model is 166 (C1 � 166); there are many combinations resources and the quality of medical services by balancing
for single-bed wards (K11 ), double-bed wards (K12 ), and bed numbers and BOR between different departments. Our
triple-bed wards (K13 ), such as (26, 40, 20), (22, 42, 20), and model may be applied in two ways. Firstly, it can be extended
(36, 20, 30). This means that W9 can provide 26 single-bed to other wards with different arrival rates and LOS distri-
wards, 40 double-bed wards, and 20 triple-bed wards; 22 bution. Secondly, our study can provide a reference for
single-bed wards, 42 double-bed wards, and 20 triple-bed dealing with the problem of hospital bed capacity to other
wards; or 36 single-bed wards, 20 double-bed wards, and 30 large general hospitals in China. Our research provides a
triple-bed wards; and so on. common framework for hospital bed allocation, so other
departments or hospitals can follow our three-stages
4. Discussion framework to realize their allocation of beds. Because the
data of each hospital and the actual situation are different
Many tertiary hospitals in China are facing the same from those of WCH, different constraints or objective
problem as WCH, with respect to the imbalance in the functions may be generated. For example, other hospitals
can use data analysis to screen the departments that need to of the hospital, management levels, equipment conditions,
allocate beds; then, they can follow the method described in scientific research capabilities, and more. The West China
Stage II to fit the functional relationship between BOR and Hospital is one of the highest-grade hospitals among the
bed. Different hospitals may have different functional re- tertiary hospitals.
lationships because of their different data. Finally, hospitals
can redesign the model with more personalized objective Data Availability
functions and constraints according to their own actual
situation. The data used to support the findings of this study are re-
stricted by the West China Hospital in order to protect
patient privacy. Data are available from West China Hospital
5. Conclusions for researchers who meet the criteria for access to confi-
We focused upon the problem of allocating beds among dential data.
different departments in a hospital. We took a large public
hospital in China, WCH, as a case study. To relieve im- Conflicts of Interest
balances in BOR between departments, we proposed a three-
The authors declare that there are no conflicts of interest
stage framework. In the first stage, we collected data and
regarding the publication of this paper.
identified departments of interest. In the second stage, we
identified the functional relationship between the number of
Acknowledgments
beds and the BOR. The third-stage MIP model provides the
best number of bed allocations for different departments. It The authors gratefully acknowledge the support from the
has proven to be a feasible method to ease the shortage of Admission Service Center of West China Hospital. This
beds. research was supported in part by the National Natural
Our research is based on real data, and hospital man- Science Foundation of China (Nos. 71532007 and 71131006)
agers can draw upon the results of this study to solve the bed and Sichuan Province Science and Technology Support
occupancy and capacity problem. The three-stage frame- Project Plan (No. 2016FZ0080).
work can help bed managers adjust the allocation of beds in a
timely and dynamic manner. This approach can be applied References
to the majority of other hospitals and may serve as a starting
point for the development of allocation models for other [1] World Health Organization, Tracking Universal Health
service industries with similar conditions, such as the al- Coverage: First Global Monitoring Report, World Health
location of beds or room types in hotels. Organization, Geneva, Switzerland, 2015, http://www.who.
Future study can consider the following two aspects: int/healthinfo/universal_health_coverage/report/2015/en/.
[2] Beijing: China Union Medical University Press. 2016, http://
since this is the initial stage of bed allocation, the strategy can
www.nhfpc.gov.cn/guihuaxxs/s10748/201708/d82fa7141696407
be extended to more departments. In addition, the practice abb4ef764f3edf095.shtml?from�groupmessage&isappinstalled�1.
will be a good reference for other large general hospitals in [3] T. J. Best, Hospital Bed Capacity Management, Dissertations &
China. More factors may be considered for inclusion in the Theses-Gradworks, Chicago, IL, USA, June 2015.
MIP model such as other ward resources (nurses and [4] D. J. Worthington, “Queueing models for hospital waiting
doctors), infectious patients, and the undesirability of lists,” Journal of the Operational Research Society, vol. 38,
mixed-sex rooms. no. 5, pp. 413–422, 1987.
[5] R. Akkerman and M. Knip, “Reallocation of beds to reduce
waiting time for cardiac surgery,” Health Care Management
Appendix Science, vol. 7, no. 2, pp. 119–126, 2004.
[6] K. Seung-Chul, H. Ira, Y. Karl K, and B. Thomas A, “Flexible
General hospitals have been divided into three levels in
bed allocation and performance in the intensive care unit,”
China, according to the hospital’s functions, tasks, facilities, Journal of Operations Management, vol. 18, no. 4, pp. 427–443,
technology, medical services, and scientific management. 2000.
Class I hospitals (the number of beds is 100 or fewer): [7] D. Kleinloog, F. B. Houckgeest, and D. Sierink, “Triage after
primary hospitals and health centers that provide pre- drug overdose: effect of the introduction of a medical psy-
vention, medical care, health care, and rehabilitation services chiatry unit on the allocation of ICU beds,” Critical Care,
directly to communities in a specific population. vol. 19, no. 1, p. P513, 2015.
Class II hospitals (101 to 500 beds): regional hospitals [8] L. V. Green, “How many hospital beds?,” British Medical
that provide comprehensive medical and health services to Journal, vol. 39, pp. 400–412, 2003.
multiple communities and undertake certain teaching and [9] A. J. M. Oerlemans, H. Wollersheim, N. V. Sluisveld,
research tasks. J. G. V. D. Hoeven, W. J. M. Dekkers, and M. Zegers, “Ra-
tioning in the intensive care unit in case of full bed occupancy:
Class III hospitals (also called tertiary hospitals; authors’
a survey among intensive care unit physicians,” BMC Anes-
hospital; more than 501 beds): regional or higher hospitals thesiology, vol. 16, no. 1, p. 25, 2015.
that provide high-level specialized medical and health ser- [10] P. Devapriya, C. T. Strömblad, M. D Bailey et al., “StratBAM: a
vices and perform higher education and scientific research discrete-event simulation model to support strategic hospital
tasks in several areas. Tertiary hospitals are further sub- bed capacity decisions,” Journal of Medical Systems, vol. 39,
divided into three grades based upon the technical strength no. 10, p. 130, 2015.
[11] J. C. Ridge, S. K. Jones, M. S. Nielsen, and A. K. Shahani, [28] J. A. Joines and S. D. Roberts, Simulation Modeling with Simio:
“Capacity planning for intensive care units,” European A Workbook, Simio LLC, Sewickley, PA, USA, 2013.
Journal of Operational Research, vol. 105, no. 2, pp. 346–355, [29] Q. Xiao, L. Luo, S.-Z. Zhao, X.-B. Ran, and Y.-B. Feng,
1998. “Online appointment scheduling for a nuclear medicine
[12] G. Romanin-Jacur and P. Facchin, “Optimal planning of a department in a Chinese hospital,” Computational and
pediatric semi-intensive care unit via simulation,” European Mathematical Methods in Medicine, vol. 2018, no. 10, pp. 1–13,
Journal of Operational Research, vol. 29, no. 2, pp. 192–198, 2018.
1987. [30] L. B. Holm, H. Lurås, and F. A. Dahl, “Improving hospital bed
[13] E. Akcali, M. J. Côté, and C. Lin, “A network flow approach to utilisation through simulation and optimisation: with appli-
optimizing hospital bed capacity decisions,” Health Care cation to a 40% increase in patient volume in a Norwegian
Management Science, vol. 9, no. 4, pp. 391–404, 2006. general hospital,” International Journal of Medical In-
[14] M. Utley, S. Gallivan, T. Treasure, and O. Valencia, “Ana- formatics, vol. 82, no. 2, pp. 80–89, 2013.
lytical methods for calculating the capacity required to operate [31] K. Schittkowski, “EASY-FIT: a software system for data fitting
an effective booked admissions policy for elective inpatient in dynamical systems,” Structural and Multidisciplinary Op-
services,” Health Care Management Science, vol. 6, no. 2, timization, vol. 23, no. 2, pp. 153–169, 2002.
pp. 97–104, 2003. [32] S. Belciug and F. Gorunescu, “Improving hospital bed oc-
[15] C. Lakshmi and S. A. Iyer, “Application of queueing theory in cupancy and resource utilization through queuing modeling
and evolutionary computation,” Journal of Biomedical In-
health care: a literature review,” Operations Research for
formatics, vol. 53, pp. 261–269, 2015.
Health Care, vol. 2, no. 1-2, pp. 25–39, 2013.
[33] J. K. Cochran and A. Bharti, “Stochastic bed balancing of an
[16] U. Uzunoglu Kocer and S. Ozkar, “M/C2/c/K queuing model
obstetrics hospital,” Health Care Management Science, vol. 9,
and optimization for a geriatric care center,” Applied Sto-
no. 1, pp. 31–45, 2006.
chastic Models in Business and Industry, vol. 31, no. 6,
[34] D. Whitley, “A genetic algorithm tutorial,” Statistics and
pp. 892–911, 2015. Computing, vol. 4, no. 2, pp. 65–85, 1994.
[17] S. Belciug and F. Gorunescu, “A hybrid genetic algorithm- [35] J. Orendi, “Health-care organisation, hospital-bed occupancy,
queuing multi-compartment model for optimizing inpatient and MRSA,” The Lancet, vol. 371, no. 9622, pp. 1401-1402,
bed occupancy and associated costs,” Artificial Intelligence in 2008.
Medicine, vol. 68, pp. 59–69, 2016.
[18] T. P. Roh, T. R. Huschka, M. J. Brown, and Y. N. Marmor,
“Data-driven simulation use to determine bed resource re-
quirements for the redesign of pre- and post-operative care
areas,” in Proceedings of the Winter Simulation Conference,
vol. 1, pp. 1168–1176, Washington, DC, USA, 2014.
[19] E. W. Hans and A. G. Leeftink, “Integral planning of operating
rooms and wards: tactical allocation of beds to reduce fluc-
tuations in bed utilization for Medisch Spectrum Twente,” in
Enschede, F. C. Smit, Ed., pp. 1–114, General Books LLC,
Memphis, TN, USA, 2015.
[20] J. T. Blake and M. W. Carter, “A goal programming approach
to strategic resource allocation in acute care hospitals,” Eu-
ropean Journal of Operational Research, vol. 140, no. 3,
pp. 541–561, 2002.
[21] M. M. Hossein, K. Sajad, and T. Mahya, “Reallocation of shafa
hospital beds in kerman using goal programming model,”
Electronic Physician, vol. 8, no. 8, pp. 2733–2737, 2016.
[22] I. Marques, M. E. Captivo, and M. Vaz Pato, “An integer
programming approach to elective surgery scheduling,” OR
Spectrum, vol. 34, no. 2, pp. 407–427, 2012.
[23] S. Y. Shin, Y. Brun, H. Balasubramanian, P. L. Henneman, and
L. J. Osterweil, “Discrete-event simulation and integer linear
programming for constraint-aware resource scheduling,”
IEEE Transactions on Systems, Man, and Cybernetics: Systems,
vol. 48, no. 9, pp. 1578–1593, 2018.
[24] R. B. Bachouch, A. Guinet, and S. Hajri-Gabouj, “An integer
linear model for hospital bed planning,” International Journal
of Production Economics, vol. 140, no. 2, pp. 833–843, 2012.
[25] C. Tontarski, “Modeling and analysis of OR nurse scheduling
using mathematical programming and simulation,” Disser-
tations & Theses—Gradworks, 2015.
[26] E. P. C. Kao and G. G. Tung, “Bed allocation in a public health
care delivery system,” Management Science, vol. 27, no. 5,
pp. 507–520, 1981.
[27] M. B. Dumas, “Simulation modeling for hospital bed plan-
ning,” Simulation, vol. 43, no. 2, pp. 69–78, 1984.
Hindawi
https://doi.org/10.1155/2019/3130527
Research Article
An Improved Sliding Window Area Method for T Wave Detection
Haixia Shang,1 Shoushui Wei ,1 Feifei Liu,2 Dingwen Wei,3 Lei Chen,4
and Chengyu Liu 2
1
School of Control Science and Engineering, Shandong University, Jinan 250061, China
2
School of Instrument Science and Engineering, Southeast University, Nanjing 210096, China
3
Department of Electronic & Electrical Engineering, Bath University, Bath BA27AY, UK
4
School of Science and Technology, Shandong University of Traditional Chinese Medicine, Jinan 250355, China
Correspondence should be addressed to Shoushui Wei; sswei@sdu.edu.cn and Chengyu Liu; chengyu@seu.edu.cn
Received 4 December 2018; Accepted 5 March 2019; Published 1 April 2019
Copyright © 2019 Haixia Shang et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Background. The T wave represents ECG repolarization, whose detection is required during myocardial ischemia, and the first
significant change in the ECG signal is being observed in the ST segment followed by changes in other waves like P wave and QRS
complex. To offer guidance in clinical diagnosis, decision-making, and daily mobile ECG monitoring, the T wave needs to be
detected firstly. Recently, the sliding area-based method has received an increasing amount of attention due to its robustness and
low computational burden. However, the parameter setting of the search window’s boundaries in this method is not adaptive.
Therefore, in this study, we proposed an improved sliding window area method with more adaptive parameter setting for T wave
detection. Methods. Firstly, k-means clustering was used in the annotated MIT QT database to generate three piecewise functions
for delineating the relationship between the RR interval and the interval from the R peak to the T wave onset and that between the
RR interval and the interval from the R peak to the T wave offset. Then, the grid search technique combined with 5-fold cross
validation was used to select the suitable parameters’ combination for the sliding window area method. Results. With respect to
onset detection in the QT database, F1 improved from 54.70% to 70.46% and 54.05% to 72.94% for the first and second
electrocardiogram (ECG) channels, respectively. For offset detection, F1 also improved in both channels as it did in the European
ST-T database. Conclusions. F1 results from the improved algorithm version were higher than those from the traditional method,
indicating a potentially useful application for the proposed method in ECG monitoring.
1. Introduction developing accurate automatic analysis algorithms for ECG

signals is critical, especially with respect to mobile ECG
Nowadays, an increase in the number of people suffering monitoring [4]. Furthermore, QRS complex have been
from heart diseases has been seen. Characterized by several widely investigated because of its highest amplitude over the
waveforms such as the P wave, QRS complex, and T wave, past decades. Up to now, there are many classical methods
electrocardiogram (ECG) becomes the most intuitive and for detecting QRS complex and most of the methods have
basic tool to diagnose heart diseases in clinical applications been listed in [5], and the classical widely-used methods are
which can provide essential physiological/pathological in- parabolic fitting [6], neural-network-based method [7], and
formation for clinical diagnoses and decision-making [1], convolutional neural network [8]. In addition, those
including important time interval information between the methods for detecting the QRS complex have shown high
onset and offset of different waves [2]. Besides, many sensitivity with positive predictivity (>99%) on the MIT-BIH
wearable monitoring devices have appeared in recent years, arrhythmia database [9], which can provide powerful sup-
which makes it possible to monitor ECG signals throughout port for other waves’ detections.
an individual’s daily life. Meanwhile, a large amount of ECG As one of three main waves of ECG, the T wave represents
data are generated daily, which is impossible for physicians ECG repolarization, and its absence or unusual shapes may
to view/diagnose each ECG signal manually [3]. Therefore, signify disruption in repolarization or another segment of the
heartbeat [10]. Additionally, T wave abnormalities are asso- 2. Methods

ciated with some heart diseases such as inverted T waves
found in other leads (other than the V1 to V4 leads), which is 2.1. Data. Records from two datasets are used. The first is the
related to an increase in cardiac deaths, and a tall or wide QRS QT database, which contains 105 15-minute two-channel
complex with an upright T wave is further suggestive of a ECG recordings with the sample rate of 250 Hz, and we
posterior infarction. Furthermore, during myocardial ische- chose it as the training and testing sets because multiple-type
mia, the first significant change in ECG signal is being ob- records from different databases are contained in this da-
served in ST-segment followed by changes in other waves like tabase. Besides, totally 43 recordings have manually anno-
P wave and QRS complex of ECG signal. Hence, detection of tated T wave onsets and 103 recordings have manually
the T wave is significant in clinical applications [11]. annotated T wave offsets. All records with annotations are
However, accurate/robust T wave detection still presents selected, and for each record, a 0.05–45 Hz low-pass zero-
challenges due to its low amplitude (usually 0.1 to 0.3 mV) as phase filter was applied for denoising before importing to
well as great variations in T waves’ morphologies [12], like our algorithm. Furthermore, there are usually 30 to 100
positive T wave, negative T wave, and biphasic T wave. representatively manually annotated discrete beats in each
Besides, most of the ischemic cases suffering from earlier annotated recording. Thus, an RR interval adjustment is also
STEMI (ST-elevation myocardial infarction) have a prom- needed before using these records because we used the
inent ST elimination or depression, which significantly af- manually annotated R peak locations. Table 1 shows the
fects the detection of the T onsets. Nowadays, various summarized annotated information of the QT database.
approaches based on different techniques have been pro- More detailed information about the annotations of this
posed for T wave detection, and those typical techniques are database can be found in the study by Laguna et al. [35].
wavelet [13, 14], mathematical model [15], support vector The second database is the European ST-T database,
machine (SVM) [16], artificial neural network (ANN) which consists 90 2-hour two-channel ECG recordings
[17–19], low-pass differentiation (LPD) [20], hidden Markov sampled at 250 Hz, and records of this database are only
model (HMM) [21, 22], partially collapsed Gibbs sample and used to test the robustness of our improved method. The
Bayesian (PCGS) [23], “wings” function [24], derivative European ST-T database is chosen because of its widely
curve [25], adaptive technique [26], computing the Trape- usages in evaluation of algorithms for analysis of ST and T
zium’s area [27], TU complex analyses [28], correlation wave changes [36, 37]. In this study, 23 recordings (only the
analysis [29], k-nearest neighbor [30], and sliding window first 5 minutes in each recording) were selected and were
area (SWA) [31]. In these aforementioned methods, the manually annotated for T wave onsets and offsets by a
wavelet-based method is robust to waveform morphological trained staff member because of loss of T wave international
variations but is sensitive to noise [13, 14]. The mathematical annotations. Table 1 also shows the detailed annotation
model method needs to build robust ECG templates, but information of this database. Besides, when choosing re-
when the waveform variations are large, building universal cords, if there were serious signal quality problems within
templates becomes difficult [15]. The SVM-based method is the first 5-minute episode, the following 5-minute episode
efficient but constructing efficient features is tough [16], and was used and a 0.05–45 Hz bandpass filter was applied for
the ANN-based method faces the drawback of high com- denoising for each record we chose before importing into
putational complexity [17]. As a comparison, the SWA the algorithm. We do not implement RR interval adjust-
method has low computational complexity which is also ment because R peaks were detected by jqrs method [21].
robust to noise and waveform morphological variations [31]. To verify the consistency of the annotations between the
In 2006, Zhang et al. first proposed the SWA method for two databases, we analyzed the time interval information
detecting T wave offsets and confirmed its efficiency in the between the T wave onset/offset and the corresponding R
QT database [31]. Subsequently, Song et al. improved this peak position for the two databases. Figure 1 shows the
method for detecting T wave onsets [32]. Afterwards, our probability density distributions of the time interval in-
team combined onsets and offsets detection for classifying formation from the two databases. As shown in Figure 1,
the morphology of the ST segment [33]. In 2017, our team we found that our manual annotations of the onset/offset
analyzed its efficiency in the QT database with a different of T wave in the European ST-T database had similar
evaluation index (F1 measure), and we found that there is probability density distributions with the annotations in
still some space for further improvement since the pa- the QT database, which indicated the effectiveness of our
rameter settings in the transitional SWA method are not annotations.
adaptive [34], and the parameters given by Zhang et al. [31]
and Song et al. [32] are empiric values and there is no
optimization step included. 2.2. Sliding Window Area (SWA) Method
Hence, in this study, an improved SWA method for both
onset and offset detections of T wave with more adaptive 2.2.1. SWA Method. SWA is an algorithm for detecting T
parameter settings is proposed. The performance of the wave onset and offset (Ton and Tend, respectively) by analyzing
improved method was compared with the traditional method, the waveform area of ECG within a sliding window [31].
and both methods were validated in two common ECG Onset/offset is detected when the area of the sliding window
databases: (1) the QT database (training and testing) and (2) reaches its maximum in a prefixed searching range. Then, we
another independent European ST-T database (only testing). show an example for explaining the method in Figures 2 and 3.
Table 1: Summary of the annotative information of the QT and the European ST-T databases.
QT database European ST-T database
Variable
Onset Offset Onset Offset
No. of recordings 43 105 23 23
No. of annotated beats 1371 3542 14337 14337
Min. Dis_qrs (ms) 52 228 60 220
Max. Dis_qrs (ms) 412 784 264 612
Mean of Dis_qrs (ms) 164 360 160 380
SD of Dis_qrs (ms) 60 71 29 47
Dis_qrs: the time interval between the T wave onset/offset and the R wave position within the current beat; SD: standard deviation. The annotations of the QT
database are taken from the database website https://www.physionet.org/physiobank/database/qtdb/doc/index.shtml.
25 25
Probability density (%)
Probability density (%)

20 20
15 15
10 10
5 5
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Time interval from QRS complex annotation (s) Time interval from QRS complex annotation (s)
Onset
Onset
Offset
Offset
(a) (b)
Figure 1: Probability density distribution of the time interval information between T wave onset/offset and R wave peak by analyzing the
annotations from the two databases: (a) QT database and (b) European ST-T database.
Searching window Searching window
Ton t+w Tend Ton t t+w Tend
(a) (b)
Figure 2: Demonstration of the SWA method for T wave onset detection.
Searching window Searching window
Ton t–w t Tend Ton t–w Tend
(a) (b)
Figure 3: Demonstration of the SWA method for detecting offsets of the T wave.
Figure 2 illustrates the detection for Ton. Firstly, with the boundaries, but the search boundaries are closely related to
location of R peak, the left and right boundaries (t1 and t2, the RR interval. As shown in Figures 2 and 3, if the interval of
respectively) of search window are determined based on the the searching window’s boundaries was set too small which
current RR interval as suggested in the study by Song et al. [32]: means that two boundary points are near the current R peak,
􏽰�� the maximum of sliding area could not be found or the
⎪
⎧
⎪ t1 � 􏼐⌈0.5 × RRi ⌉ + Ri + 0.08􏼑s,
⎪
⎪ detected onset/offset of T wave are nearer to the R peak.
⎪
⎨ t2 � ⌊0.15 × RRi 􏼁⌋ + Ri + 0.12􏼁s, if RRi < 0.88 s,
􏽰�� (1) These issues affect detection accuracy, which results in de-
⎪
⎪ t1 � 􏼐⌈0.5 × RRi ⌉ + Ri + 0.1􏼑s, tection error and vice versa.
⎪
⎪
⎪
⎩ t � R + 0.32􏼁s, In the traditional SWA method, there are two piecewise
2 i if RRi ≥ 0.88 s,
functions with predefined parameter settings. In order to
where RRi is the ith RR interval and Ri is the ith position of R more accurately model the relationships between RR interval
peak. and the searching boundaries in this study, we performed a
The waveform area (area of onset denoted as: Ao) within k-means clustering analysis between RR intervals and RTon
the fixed sliding window 􏼂 t t + w 􏼃 was calculated using the (RTon denotes the time interval between the R peak and T
following formula: wave onset) as well as the relationship between the RR
t+w intervals and RToff (RToff the time interval between the R
Ao � 􏽘 􏼐sj − sk 􏼑, (2) peak and T wave offset), which is implemented by means of
j�t
the k-means function in Matlab. The scatter plots with the
where w � 0.12 s (by default), which is the window width, t optimal k-means clustering (k � 3) are shown in Figure 4
stretches from t1 to t2, sj is the waveform amplitude at the jth [38], and k is determined by combining the results of
sample point, and sk is the local average amplitude (using a clustering and the computational complexity of parameters’
smoothing window of p � 0.016 s by default), which is settings as well as the adaptiveness of the algorithm. Then,
defined according to the following equation: the two relationships (between RR intervals and RTon , and
t+p
between RR intervals and RToff ) are obtained using the
1 following equations:
Sk � 􏽘 S. (3)
2p + 1 j�t−p j case 1 : RR < 0.76 s, 0.05 s < RTon < 0.25 s,
As shown in Figure 2, when t � Ton, Ao reaches its case 2 : 0.76 s ≤ RR < 1.13 s, 0.05 s < RTon < 0.35 s,
maximum value. case 3 : RR ≥ 1.13 s, 0.05 s < RTon < 0.45 s,
Figure 3 illustrates the Tend detection. At first, with the (6)
case 1 : RR < 0.72 s, 0.2 s < RToff < 0.45 s,
location of the R peak, the left and right boundaries (t3 and t4,
respectively) of the search window are determined based on the case 2 : 0.72 s ≤ RR < 1.1 s, 0.2 s < RToff < 0.6 s,
current RR interval as suggested in a study by Zhang et al. [31]: case 3 : RR ≥ 1.1 s, 0.2 s < RToff < 0.8 s.
⎪
⎧
⎪ t3 � ⌊0.15 × RRi ⌋ + Ri + 0.148􏼁s,
⎪
⎪ Thus, the three piecewise functions for determining the
⎪
⎪
⎪
⎨ t4 � ⌈0.7 × RRi ⌉ + Ri − 0.036􏼁s, if RRi < 0.88 s, search boundaries for T wave onset and offset detections
⎪ (4) were obtained with the parameters presented in Table 2:
⎪
⎪ t3 � Ri + 0.28􏼁s,
⎪
⎪ 􏽰��
⎪
⎪ ⎪
⎧ t1 �􏼐Ri +⌈ald × RRi ⌉ + 0.02􏼑􏼑s,
⎩ ⎪
⎪
t4 � ⌈0.2 × RRi ⌉ + Ri + 0.404􏼁s, if RRi ≥ 0.88 s. ⎪
⎪
⎪ 􏽰��
⎪
⎪ t2 � 􏼐Ri +⌈alu × RRi ⌉ + 0.16􏼑s, if RRi < 0.76 s,
The waveform area (area of ends denoted as: Ae) within ⎪
⎪
⎪
⎪ 􏽰 ��
the fixed sliding window 􏼂 t − w t 􏼃 was then calculated ⎪
⎪
⎨ t1 � 􏼐Ri +⌊ard × RRi ⌋ + 0.04􏼑s,
according to the following formula: ⎪ 􏽰��
⎪
⎪ t2 � 􏼐Ri +⌊aru × RRi ⌋ + 0.24􏼑s, if 0.76 s ≤ RRi < 1.13 s,
t ⎪
⎪
⎪
⎪ 􏽰��
Ae � 􏽘 􏼐sj − sk 􏼑, (5) ⎪
⎪
⎪
⎪ t1 � 􏼐Ri +⌈amd × RRi ⌉ + 0.04􏼑s,
j�t−w ⎪
⎪
⎪
⎩ t � 􏼐R +⌈amu × 􏽰��
2 i RRi ⌉ + 0.4􏼑s, if RRi ≥ 1.13 s,
where w � 0.128 s (by default), t is from t3 to t4, and sj and sk
have been defined in equation (2). As shown in Figure 3, (7)
when t � Tend, Ae reaches its maximum value. As for the
difference between Figures 2 and 3 is the direction to cal- ⎪
⎧
⎪ t3 � Ri +⌈ald × RRi ⌉ + 0.18􏼁s,
⎪
⎪
culate the sliding area. ⎪
⎪ t4 � Ri +⌈alu × RRi ⌉ + 0.3􏼁s, if RRi < 0.72 s,
⎪
⎪
In addition, Algorithm 1 shows the description of the ⎪
⎪
⎨ t3 � Ri +⌊ard × RRi ⌋ + 0.18􏼁s,
traditional SWA algorithm and more details to which the ⎪
algorithm proof can refer [31]. ⎪
⎪ t4 � Ri +⌊aru × RRi ⌋ + 0.4􏼁s, if 0.72 s ≤ RRi < 1.1 s,
⎪
⎪
⎪
⎪ � Ri +⌈amd × RRi ⌉ + 0.18􏼁s,
⎪ t3
⎪
⎪
⎩
t4 � Ri +⌈amu × RRi ⌉ + 0.48􏼁, s if RRi ≥ 1.1 s.
2.2.2. Improved SWA Method. One key issue with respect to
the SWA method is to accurately determine the search (8)
0.5
0.4
0.3
RTon
0.2
0.1
0
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
RR interval (s)
Cluster 1 Cluster 3
Cluster 2 Centroids
(a)
1
0.8
0.6
RToff
0.4
0.2
0
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
RR interval (s)
Cluster 1 Cluster 3
Cluster 2 Centroids
(b)
Figure 4: Clustering results for T wave feature points: (a) clustering information of T wave onsets; (b) clustering information of T wave
offsets.
Input: ECG signal S(t), R peak locations, R peak numbers N, sliding window width w, smoothing factor p, and morphology
predefined factor r
Output: T wave onset locations
Calculation:
(1) Calculate t1 and t2 and construct sliding window
window � S(t1 : t2 )
(2) Smooth signal and calculate sliding area for each point i inside [t1, t2]
Area(i) � sum(S(i : (i + w)) − sum(S((i − p) : (i + p)))/(p ∗ 2 + 1))
(3) T wave morphology classification
if r � “p”‖r � “pn”‖r � “bm”‖r � “”
calculate [k1 , l1 ] � max(Area)
end if
if r � “n”‖r � “np”‖r � “”
calculate [k2 , l2 ] � max(−Area)
end if
(4) Get k
k � min(k1 , k2 )
Output: T wave onset � R peak + k.
ALGORITHM 1: Traditional SWA algorithm (T wave onset detection).
Then, the grid search was used to determine the best well as using a 5-fold cross-validation. Then, we stored the F1
combination of parameters in equations (7) and (8), which measure of one loop and started another loop. Through all
was implemented by for loop. In a loop, we changed the loops, we traversed all of the combinations of parameters
value of one parameter at a time, kept the other parameters referred to in Table 2. After comparing the results, the
unchanged, and applied the algorithm in the QT database as combinations of parameters with the highest F1 measure
were chosen. The best parameters’ combinations for T wave onset detection. F1 improved from 41.02% to 84.13% and
onsets are listed: ald � 0.4, alu � 0.2, ard � 0.4, aru � 0.4, 44.33% to 87.62% for two ECG channels, respectively. The
amd � 0.3, and amu � 0.0 and for T wave ends are listed: mean detection errors significantly decreased from 19.52 ms
ald � 0.2, alu � 0.1, ard � 0.2, ard � 0.1, aru � 0.0, amd � 0.0, to 7.04 ms and 36.27 ms to 6.35 ms for two ECG channels,
and amu � 0.1. The improved SWA method can be sum- respectively. Performance improvements in offset detection
marized as a block diagram in Figure 5. were small but convincing F1 improved from 98.83% to
99.57% and 91.76% to 98.29% for two ECG channels, re-
spectively. However, the mean detection errors for T wave
2.3. Evaluation Method. Detections for true and false positives offset detection slightly increased (not significant) when
(TP and FP, respectively) and false negative (FN) were de- performing the improved method.
termined with a threshold of 100 ms. In this study, indices like
sensitivity (Se), positive precision (P+), and F1 measurement
were selected as evaluation indices [39, 40] with the following 4. Discussion
definitions: Se � TP/(TP + FN), P+ � TP/(TP + FN), and As seen from Tables 3 and 4, both T wave onset and offset
F1 � (TP × 2)/(TP × 2 + FN + FP). F1 measure is selected detection of the new proposed method reported better
other than accuracy since F1 measure is the weighted average of performances (F1 measure) than the traditional method,
precision and recall which satisfies our asymmetric datasets suggesting that applying the clustering technique in the
where values of false positive and false negatives are not the SWA method for deciding searching boundaries is helpful to
same. enhance detection accuracy. In addition, clustering is a
statistical-based technique, which can be used to determine
3. Results whether the independent part of a population belongs to
different groups by comparing quantitative multiple features
Figure 6 shows the detection examples of the proposed [38]. Besides, we noted that, for the T wave offset detections,
method, compared with the traditional methods, Zhang’s neither the traditional SWA nor the improved version re-
method for T wave offset detection [31] and Song’s method ported the better performance than the detection of T wave
for T wave onset detection [32]. Figure 6(a) shows the onsets. One possible explanation is that Zhang proposed this
inverted T wave detections, Figure 6(b) shows the biphasic T method originally to detect T wave offsets not T wave onsets
wave detections, and Figure 6(c) shows the normal T wave and proved its mathematical rationality for T wave offsets.
detections. From Figure 6, T wave offset detections get better Another possible explanation is that the clustering method
results than T wave onset detections. And, our method got for determining the search boundaries is a statistical-based
obviously better results when it is applied in T wave onsets technique. Therefore, the accuracy of the clustering results is
detections. related to the data amount. However, the annotated T wave
onsets in the QT database are far less than the annotated T
3.1. Results from the QT Database. We firstly tested the wave offsets (1371 versus 3452). Thus, the relationship found
performance of the improved SWA method on the QT by clustering analysis between the RR interval and RTon is
database. The traditional SWA methods (Song’s method [32] not that strong (Figure 4(a)) compared to the relationship
and Zhang’s method [31]) were used as comparators. between the RR interval and RToff (Figure 4(b)). Moreover,
Table 3 shows the results of onset and offset detections in the significant difference between the improved and tradi-
the QT database. Both of the two channels signals (first and tional SWAs indicates that the improved version can more
second channels) were tested. From Table 3, we found the extensively and adaptively determine the search window’s
improved SWA method significantly enhanced detection boundaries by using the k-means clustering based on the QT
accuracies for both onset and offset detections. For onset database and grid search strategy. However, the traditional
detection, F1 improved from 54.70% to 70.46% and 54.05% SWA only used predefined parameters and did not give out
to 72.94% for two ECG channels, respectively. For offset any detailed explanations.
detection, F1 improved from 87.83% to 93.73% and 86.73% Another difference between Zhang’s [31] and our results
to 94.75% for two ECG channels, respectively. In addition, was observed when using the QT database for validation,
detection errors were also analyzed. As expected, the im- Zhang’s study chose the better result from the outputs of the
proved SWA method indicated smaller detection errors than two ECG channels [31]. In order to compare our results with
the traditional method except for a slight increase in the those from Zhang, we also calculated smaller errors from the
offset detection from the second channel (traditional results of two ECG channels. The comparable results are
0.027 ± 31.85 ms versus improved 2.45 ± 33.98 ms). How- summarized in Table 5. The mean detection errors are
ever, it is worthwhile to note that all Se, P+, and F1 indices similar between Zhang’s and our results. We also noted that
increased from ∼86% to ∼94%. the standard deviation of detection errors was 25.82 ms for
our method and 21.19 ms for the traditional SWA. Both of
them were smaller than the acceptable threshold (30.6 ms)
3.2. Results from the European ST-T Database. Table 4 shows proposed by the common standards in Electrocardiography
the results of onset and offset detections in the European ST- Working Party [41].
T database. The improvements after using the improved Table 5 also summarizes comparable results from other
method were more significant when performing T wave studies. The wavelet-based method reported a mean error of
Input QT database signal Initializing parameters
Clustering result learning
Setting parameters’ range
Fivefold cross-validation Grid search
Comparing F1 measure
Input European ST-T

Improved method
database signal
Results of two databases
Figure 5: The block diagram of the proposed method for delineating the T wave onset/offset.
1
Voltage (mV)
0.5
–0.5
–1
0.5 1 1.5 2 2.5 3 3.5
Time (s)
ECG T onset-proposed method
R peak T offset-Zhang’s method
T offset-proposed method T onset-Song’s method
(a)
1
Voltage (mV)
0.5
–0.5
–1
0.5 1 1.5 2 2.5 3 3.5
Time (s)
(b)
Figure 6: Continued.
Voltage (mV)
0.5
–0.5
–1
0.5 1 1.5 2 2.5 3 3.5
Time (s)
(c)
Figure 6: T wave detection examples. The solid points marked ( ) are R peaks; the solid points marked ( and ) are results of our method; the
hollow circles marked ( and ) are results of the traditional method; the shadow areas are accepted as TP cases. (a) e0107; (b) e0111; (c)
e0118.
Table 2: Information of parameters when detecting the T wave.

T onset T offset
Parameters
IV CS CR IV CS CR
ald/alu/ard 0.1 0.1∼0.4
0.1 0.1∼0.4
aru/amd 0.1 0.1
0.0 0.0∼0.4
amu 0.0 0.0∼0.4
IV: initialized value; CS: change step; CR: change range.
Table 3: Results of T wave detection in the QT database.

Detection Channel Method Se (%) P+ (%) F1 (%) Error mean ± SD (ms)
Traditional SWA [32] 54.70 54.70 54.70 −30.2 ± 40.75
First
Improved SWA 70.46 70.46 70.46 7.3 ± 53.12
Onset
Traditional SWA [32] 54.05 54.05 54.05 −36.27 ± 43.29
Second
Improved SWA 72.94 72.94 72.94 6.35 ± 53.78
Traditional SWA [31] 87.83 87.83 87.83 −2.57 ± 30.08
First
Improved SWA 93.93 93.93 93.93 1.19 ± 33.59
Offset
Traditional SWA [31] 86.73 86.73 86.73 0.027 ± 31.85
Second
Improved SWA 94.75 94.75 94.75 2.45 ± 33.98
Table 4: Results of T wave detection in the European ST-T database.

Detection Channel Method Se (%) P+ (%) F1 (%) Error mean ± SD (ms)
Traditional SWA [32] 41.02 41.02 41.02 19.52 ± 31.89
First
Improved SWA 84.13 84.13 84.13 27.87 ± 44.22
Onset
Traditional SWA [32] 44.33 44.32 44.33 −36.27 ± 43.29
Second
Improved SWA 87.62 87.61 87.62 215.20 ± 33.54
Traditional SWA [31] 98.80 98.86 98.83 22.2 ± 22.70
First
Improved SWA 99.50 99.65 99.57 26.94 ± 20.98
Offset
Traditional SWA [31] 91.80 91.71 91.76 21.07 ± 26.31
Second
Improved SWA 98.29 98.28 98.29 24.54 ± 25.52
1.6 ms [13, 14]; the low-pass differentiation-based method based method reported 2.8 ms [20]. The TU complex
gave a relative large mean error of 13.5 ms [20], while the analysis gave a minimum detection mean error of 0.8 ms but
hidden Markov model-based method reported a mean error did not include the corresponding Se and P+ results [28]. In
of 5 ms [21, 22]. Furthermore, the partially collapsed Gibbs addition, methods proposed by Mahsa with linear and
sample reported 4.3 ms [23], and the k-nearest neighbor- nonlinear phase observation to detect fiducial points are also
Table 5: Comparable detection results of T wave offset in the QT database.

Methods Annotations Se (%) P+ (%) Mean ± SD (ms)
Improved SWA 3542 98.5 98.5 1.21 ± 25.82
Traditional SWA [31] 3542 95.5 95.5 −1.12 ± 21.19
Wavelet-based [13, 14] 3542 99.77 97.79 −1.6 ± 18.1
Low-pass differentiation-based [20] 3542 99.00 97.74 13.5 ± 27.0
Hidden Markov model-based [21, 22] 3542 NA NA −5 ± 14
Partially collapsed Gibbs sample [23] 3403 99.81 98.97 4.3 ± 20.8
k-nearest neighbor-based [30] 30 records NA NA 2.8 ± 18.6
TU complex analysis [28] 3528 92.60 NA 0.8 ± 30.3
Neural network and fixed-size least-squares SVM [19] 3542 NA NA −3.0 ± 16.9
L.EKF25 [42] 11 ± 39
10 records NA NA
N.L.EKF25 [42] 4 ± 23
L.EKF25 [42] −17 ± 30
15 records NA NA
N.L.EKF25 [42] −21 ± 19
NA: not available; L.EKF25: linear Kalman filter; N.L.EKF25: nonlinear Kalman filter.
listed as comparative method [42], and two parts of QT

3000
database including normal sinus rhythm and arrhythmia
Number of beats
database are used when evaluating extraction of fiducial 2500
points and the nonlinear observation has more smaller 2000
deviations 23 ms for the first database and 19 ms for the 1500

second database. 1000
The potential issues existing in the above studies consist 500
of two main points: (1) the time tolerance for determining 0
–100 –80 –60 –40 –20 0 20 40 60 80 100
true positive detection was not clear and (2) training and Interval of error (ms)
testing were both performed in the QT database, but we used
TR result
the European ST-T database as the independent testing IM result
dataset.
Besides, as deep-learning technology improves, more Figure 7: The cumulative line chart of error (T offsets in the QT
and more methods based on this technique have been database). TR result represents the traditional SWA method result,
proposed to detect ECG feature points; for instance, a re- and IM result represents the improved SWA method result.
cently proposed method using neural network and fixed-size
least-squares SVM to detect T wave end reported it is a annotations. Besides, we only combine the data statistic and
minimum detection mean error of −3 ms in the QT database; data mining technique to changing the parameters of tradi-
a QRS complex detection by using two-level convolutional tional SWA method. In our following work, more records with
neural network [8] reported its sensitivity of 99.77% in the authoritative annotations will be used to test the robustness of
MIT-BIH AR database. When using deep-learning tech- the combination of parameters we obtained in this study.
nique, a great amount of data is needed, and in T wave
detection, the annotated T wave ends are limited but a 5. Conclusion
meaningful strategy was proposed in [19], which is to use
different strategies for selecting different training sets such as In this paper, an improved sliding window area method for
random selection and k-means. But, we just proposed an detecting T wave onset and offset was proposed. The main
idea that is to use one independent database (QT dataset) as contribution/novelty was for application of the data statistic
the training set and testing set and another independent and data mining technique: (1) k-means clustering for the
database (records from the European ST-T database an- setting of search boundaries and (2) grid search strategy to
notated by a trained staff) as the testing set. optimize the parameters. Experiments performed in the QT
In order to illustrate difference between error and F1 database and the European ST-T database demonstrated the
measure, we did statistical analysis of error. And, Figure 7 improved method’s better performance.
gives the cumulative line chart of error (denoted as CLCE) of
T wave offsets in the QT database which explains our method Data Availability
got more true positive beats than the traditional method
inside our time tolerance. The CLCE of T wave onsets in the The data used to support the findings of this study are
QT database and CLCE of T wave in the European ST-T available from the corresponding author upon request.
database also have the same regularities of distribution as it
did in the T wave offsets in the QT database. Conflicts of Interest
Moreover, the limitation of our study is that the anno-
tations of the European ST-T database are only done by a The authors declare no conflicts of interest regarding the
trained staff member which may result in nonauthoritative publication of this work.
Acknowledgments [14] J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, and

P. Laguna, “A wavelet-based ECG delineator: evaluation on
This work was supported by the National Natural Science standard databases,” IEEE Transactions on Biomedical Engi-
Foundation of China under Grant no. 81871444, Shandong neering, vol. 51, no. 4, pp. 570–581, 2004.
Province Key Research and Development Plan under Grant [15] J. P. V. Madeiro, W. B. Nicolson, P. C. Cortez et al., “New
no. 2018GSF118133, and Jiangsu Province Primary Research approach for T-wave peak detection and T-wave end location
and Development Plan under Grant no. BE2017735. The in 12-lead paced ECG signals based on a mathematical
authors thank the support from the Southeast-Lenovo model,” Medical Engineering & Physics, vol. 35, no. 8,
Wearable Heart-Sleep-Emotion Intelligent Monitoring Lab. pp. 1105–1115, 2013.
[16] S. Mehta, N. Lingayat, and S. Sanghvi, “Detection and de-
lineation of P and T waves in 12-lead electrocardiograms,”
References Expert Systems, vol. 26, no. 1, pp. 125–143, 2009.
[17] G. Vijaya, V. Kumar, and H. K. Verma, “Artificial neural
[1] J. H. O’Keefe, S. C. Hammill, and M. Freed, The Complete network based wave complex detection in electrocardio-
Guide to ECGs, Physicians’ Press, Boston, MA, USA, 1997. grams,” International Journal of Systems Science, vol. 28, no. 2,
[2] J. P. V. Madeiro, P. C. Cortez, J. A. L. Marques, pp. 125–132, 1997.
C. R. V. Seisdedos, and C. R. M. R. Sobrinho, “An innovative [18] A. A. S. León et al., “Neural network approach for T-wave end
approach of QRS segmentation based on first-derivative, detection: a comparison of architectures,” in Proceedings of
Hilbert and Wavelet Transforms,” Medical Engineering & the Computing in Cardiology Conference, Vancouver, BC,
Physics, vol. 34, no. 9, pp. 1236–1246, 2012. Canada, September 2016.
[3] E. J. Topol, “The big medical data miss: challenges in estab- [19] A. A. Suárez-León, C. Varon, R. Willems, S. Van Huffel, and
lishing an open medical resource,” Nature Reviews Genetics, C. R. Vázquez-Seisdedos, “T-wave end detection using neural
vol. 16, no. 5, pp. 253-254, 2015. networks and Support Vector Machines,” Computers in Bi-
[4] A. Page, O. Kocabas, T. Soyata, M. Aktas, and J.-P. Couderc, ology and Medicine, vol. 96, pp. 116–127, 2018.
“Cloud-based privacy-preserving remote ECG monitoring [20] P. Laguna, R. Jané, and P. Caminal, “Automatic detection of
and surveillance,” Annals of Noninvasive Electrocardiology, wave boundaries in multilead ECG signals: validation with the
vol. 20, no. 4, pp. 328–337, 2015. CSE database,” Computers and Biomedical Research, vol. 27,
[5] B.-U. Kohler, C. Hennig, and R. Orglmeister, “The principles no. 1, pp. 45–60, 1994.
of software QRS detection,” IEEE Engineering in Medicine and [21] M. Akhbari, M. B. Shamsollahi, O. Sayadi, A. A. Armoundas,
Biology Magazine, vol. 21, no. 1, pp. 42–57, 2002. and C. Jutten, “ECG segmentation and fiducial point ex-
[6] A. I. Manriquez and Q. Zhang, “An algorithm for QRS onset traction using multi hidden Markov model,” Computers in
and offset detection in single lead electrocardiogram records,” Biology and Medicine, vol. 79, pp. 21–29, 2016.
in Proceedings of the 2007 29th Annual International Con- [22] D. A. Coast, R. M. Stern, G. G. Cano, and S. A. Briller, “An
ference of the IEEE Engineering in Medicine and Biology approach to cardiac arrhythmia analysis using hidden Markov
Society, Lyon, France, August 2007. models,” IEEE Transactions on Biomedical Engineering,
[7] Q. Xue, Y. H. Hu, and W. J. Tompkins, “Neural-network- vol. 37, no. 9, pp. 826–836, 1990.
based adaptive matched filtering for QRS detection,” IEEE [23] C. Lin, C. Mailhes, and J. Y. Tourneret, “P- and T-wave
Transactions on Biomedical Engineering, vol. 39, no. 4, delineation in ECG signals using a Bayesian approach and a
pp. 317–329, 1992. partially collapsed Gibbs sampler,” IEEE Transactions on
[8] Y. Xiang, Z. Lin, and J. Meng, “Automatic QRS complex Biomedical Engineering, vol. 57, no. 12, pp. 2840–2849, 2010.
detection using two-level convolutional neural network,” [24] M. Zarrini and A. Sadr, “A real-time algorithm to detect
BioMedical Engineering OnLine, vol. 17, no. 1, p. 13, 2018. inverted and symmetrical T-wave,” in Proceedings of the
[9] F. Liu, C. Liu, X. Jiang et al., “Performance analysis of ten International Conference on Computer and Electrical Engi-
common QRS detectors on different ECG application cases,” neering (ICCEE 2009), Dubai, UAE, December 2009.
Journal of Healthcare Engineering, vol. 2018, Article ID [25] Y. N. Singh and P. Gupta, “An efficient and robust technique
9050812, 8 pages, 2018. of T wave delineation in electrocardiogram,” in Proceedings of
[10] E. B. Hanna and D. L. Glancy, “ST-segment depression and the International Conference on Bio-Inspired Systems and
T-wave inversion: classification, differential diagnosis, and Signal Processing (Biosignals 2009), Porto, Portugal, January
caveats,” Cleveland Clinic Journal of Medicine, vol. 78, no. 6, 2009.
pp. 404–414, 2011. [26] N. Bayasi et al., “Adaptive technique for P and T wave de-
[11] J. Tikkanen, T. Kenttä, K. Porthan, H. V. Huikuri, and lineation in electrocardiogram signals,” in Proceedings of the
M. J. Junttila, “Electrocardiographic T wave abnormalities and 2014 36th Annual International Conference of the IEEE En-
the risk of sudden cardiac death: the Finnish perspective,” gineering in Medicine and Biology Society, Chicago, IL, USA,
Annals of Noninvasive Electrocardiology, vol. 20, no. 6, August 2014.
pp. 526–533, 2015. [27] C. R. Vázquez-Seisdedos, J. Neto, E. J. Marañón Reyes,
[12] S. A. Immanuel, A. Sadrieh, M. Baumert et al., “T-wave A. Klautau, and R. C. Limão de Oliveira, “New approach for
morphology can distinguish healthy controls from LQTS T-wave end detection on electrocardiogram: performance in
patients,” Physiological Measurement, vol. 37, no. 9, noisy conditions,” BioMedical Engineering OnLine, vol. 10,
pp. 1456–1473, 2016. no. 1, p. 77, 2011.
[13] A. Ghaffari, M. R. Homaeinezhad, M. Akraminia, M. Atarod, [28] J. A. Vila, Y. Gang, J. M. R. Presedo, M. Fernandez-Delgado,
and M. Daevaeiha, “A robust wavelet-based multi-lead S. Barro, and M. Malik, “A new approach for TU complex
Electrocardiogram delineation algorithm,” Medical Engi- characterization,” IEEE Transactions on Biomedical Engi-
neering & Physics, vol. 31, no. 10, pp. 1219–1227, 2009. neering, vol. 47, no. 6, pp. 764–772, 2000.
[29] M. R. Homaeinezhad, M. Erfanianmoshiri-Nejad, and

H. Naseri, “A correlation analysis-based detection and de-
lineation of ECG characteristic events using template wave-
forms extracted by ensemble averaging of clustered heart
cycles,” Computers in Biology and Medicine, vol. 44, no. 1,
pp. 66–75, 2014.
[30] I. Saini, D. Singh, and A. Khosla, “K-nearest neighbour-based
algorithm for P- and T-waves detection and delineation,”
Journal of Medical Engineering & Technology, vol. 38, no. 3,
pp. 115–124, 2014.
[31] Q. Zhang, A. I Manriquez, C Médigue, Y Papelier, and
M Sorine, “An algorithm for robust and efficient location of
T-wave ends in electrocardiograms,” IEEE Transactions on
Biomedical Engineering, vol. 53, no. 1, pp. 2544–52, 2006.
[32] J. Song, H. Yan, Z. Xiao, X. Yang, and X. Zhang, “A robust and
efficient algorithm for st-t complex detection in electrocar-
diograms,” Journal of Mechanics in Medicine and Biology,
vol. 11, no. 5, pp. 1103–1111, 2011.
[33] M. Xu, S. Wei, X. Qin, Y. Zhang, and C. Liu, “Rule-based
method for morphological classification of ST segment in
ECG signals,” Journal of Medical and Biological Engineering,
vol. 35, no. 6, pp. 816–823, 2015.
[34] H. Shang et al., “Performance evaluation for the sliding area-
based T wave detection method on the QT database,” in
Proceedings of the Chinese Automation Congress, Jinan, China,
October 2017.
[35] P. Laguna et al., “A database for evaluation of algorithms for
measurement of QT and other waveform intervals in the
ECG,” in Proceedings of the Computers in Cardiology, 1997,
Lund, Sweden, September 1997.
[36] G. B. Moody and PhysioNet, “Research resource for complex
physiologic signals,” Electrocardiology, vol. 29, pp. 1–88, 2009.
[37] A. Taddei, G. Distante, M. Emdin et al., “The European ST-T
database: standard for evaluating systems for the analysis of
ST-T changes in ambulatory electrocardiography,” European
Heart Journal, vol. 13, no. 9, pp. 1164–1172, 1992.
[38] A. K. Jain, “Data clustering: 50 years beyond K-means ☆,”
Pattern Recognition Letters, vol. 31, no. 8, pp. 651–666, 2008.
[39] S. Daskalaki, I. Kopanas, and N. Avouris, “Evaluation of
classifiers for an uneven class distribution problem,” Applied
Artificial Intelligence, vol. 20, no. 5, pp. 381–417, 2006.
[40] C. Liu, P. Li, C. Di Maria, L. Zhao, H. Zhang, and Z. Chen, “A
multi-step method with signal quality assessment and fine-
tuning procedure to locate maternal and fetal QRS complexes
from abdominal ECG recordings,” Physiological Measure-
ment, vol. 35, no. 8, pp. 1665–1683, 2014.
[41] J. Willems, “Recommendations for measurement standards in
quantitative electrocardiography,” European Heart Journal,
vol. 6, no. 10, pp. 815–825, 1985.
[42] M. Akhbari, M. B. Shamsollahi, C. Jutten, A. A. Armoundas,
and O. Sayadi, “ECG denoising and fiducial point extraction
using an extended Kalman filtering framework with linear and
nonlinear phase observations,” Physiological Measurement,
vol. 37, no. 2, pp. 203–226, 2016.
Hindawi
https://doi.org/10.1155/2019/6509357
Review Article
A Technical Review of Convolutional Neural Network-Based
Mammographic Breast Cancer Diagnosis
Lian Zou,1,2,3 Shaode Yu ,1,4 Tiebao Meng,5 Zhicheng Zhang,1,2 Xiaokun Liang,1,2,6
and Yaoqin Xie 1
1
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
2
Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen, China
3
Cancer Center of Sichuan Provincial People’s Hospital, Chengdu, China
4
Department of Radiation Oncology, The University of Texas Southwestern Medical Center, Dallas, TX, USA
5
Department of Medical Imaging, Sun Yat-sen University Cancer Center, Guangzhou, China
6
Medical Physics Division in the Department of Radiation Oncology, Stanford University, Palo Alto, CA, USA
Correspondence should be addressed to Yaoqin Xie; yq.xie@siat.ac.cn
Received 14 January 2019; Accepted 25 February 2019; Published 25 March 2019
Copyright © 2019 Lian Zou et al. This is an open access article distributed under the Creative Commons Attribution License,
This study reviews the technique of convolutional neural network (CNN) applied in a specific field of mammographic breast
cancer diagnosis (MBCD). It aims to provide several clues on how to use CNN for related tasks. MBCD is a long-standing
problem, and massive computer-aided diagnosis models have been proposed. The models of CNN-based MBCD can be broadly
categorized into three groups. One is to design shallow or to modify existing models to decrease the time cost as well as the number
of instances for training; another is to make the best use of a pretrained CNN by transfer learning and fine-tuning; the third is to
take advantage of CNN models for feature extraction, and the differentiation of malignant lesions from benign ones is fulfilled by
using machine learning classifiers. This study enrolls peer-reviewed journal publications and presents technical details and pros
and cons of each model. Furthermore, the findings, challenges and limitations are summarized and some clues on the future work
are also given. Conclusively, CNN-based MBCD is at its early stage, and there is still a long way ahead in achieving the ultimate
goal of using deep learning tools to facilitate clinical practice. This review benefits scientific researchers, industrial engineers, and
those who are devoted to intelligent cancer diagnosis.
1. Introduction Mammography performs as a routine tool for breast

cancer screening. It enables high-resolution perception of
Breast cancer threatens women’s life worldwide. In the the internal anatomy of breast and helps the diagnosis of
United States, it might cause an estimation of 0.25 million suspicious lesions [6]. Screening mammography scans the
new cases of invasive breast cancer, 0.06 million new cases of breast from the craniocaudal view and mediolateral oblique
noninvasive breast cancer, and 0.04 million deaths in 2016 view, while diagnostic mammography acquires more images
[1]. This disease dramatically increases the health burden on when symptoms, such as architecture changes and abnormal
those developing and underdeveloped countries [2]. Sub- findings, are found on screening mammographic images. To
stantial clinical trial indicates that early detection and di- date, screen film mammography (FM) has been the reference
agnosis of breast cancer can provide patients with more standard for use in breast cancer screening programs, while
flexible treatment options and improved life quality and due to the demands of higher spatial resolution, digital
survivability [3]. Therefore, more and more attention is mammography (DM) has been widely accepted. General
being paid to related fields, such as novel imaging modalities rules exist for mammographic image interpretation. How-
of ultrasound tomography [4] and breast tomography [5]. ever, errors are unavoidable in clinic, and reasons are
manifold. Above all, the difference of perceived visual ap- This paper also presents a review. It is dedicated to the
pearance between malignant and benign lesions is unclear technique of CNN applied in a specific application of
and consequently, how to quantify breast lesions with dis- MBCD, and it aims to provide clues on how to use CNN in
criminative features is full of challenges. Moreover, it is still intelligent diagnosis. The contributions of this review are
difficult to estimate the disease risk because of limited in- summarized as follows. At first, this study is restricted to
formation and thus, healthy people might be turned into peer-reviewed journal publications and consequently,
patients. Besides, work overload and fatigue further cause technical details and pros and cons of each model can be
misinterpretation and overdiagnosis. Unfortunately, it is delivered. Furthermore, based on how to use CNN tech-
found that more than 70% of suggested biopsies are with niques, the MBCD models are broadly categorized into three
benign outcomes during the diagnosis phase [7]. groups. One is to design shallow models or to modify
Computer-aided models for mammographic breast existing models for decreased time cost and medical in-
cancer diagnosis (MBCD) have been explored for over thirty stances for training; another is to make the best use of a
years [8, 9]. It supports the decision making and helps the pretrained CNN model by transfer learning and parameter
differentiation between malignant and benign lesions by fine-tuning; and the third is to take advantage of CNN
providing additional information. Due to the facilitation of models for feature extraction, while the differentiation be-
MBCD models, the diagnostic performance is enhanced tween malignant and benign lesions is based on machine
regarding both sensitivity and specificity [10] and un- learning classifiers. At last, findings, challenges, and limi-
necessary examinations can be reduced in a cost-effective tations are summarized, and some clues on the future work
manner. It further benefits biopsy recommendations, follow- are also given.
up treatments, and prognosis analysis. From a technical The remainder of this paper is structured as follows.
perspective, major MBCD models are consisted of feature Section 2 describes basic concepts regarding computer-aided
extraction and lesion malignancy prediction. The former diagnosis (CAD) and transfer learning. Section 3 reviews
quantifies lesions with discriminative features and the latter CNN-based MBCD techniques, including the search strat-
builds the relationship between the features and its label, egy of the literature and technical details of involved models.
benign or malignant. Massive studies have devoted to the And then, findings, challenges, and future focus are sum-
investigation of breast cancer diagnosis, ranging from using marized in Section 4. In the end, Section 5 concludes this
different modalities [11–13], to the analysis of subtle signs review.
[14, 15] and to various technique exploration [16, 17]. Be-
cause of the easy accessibility of high-performance com- 2. Basic Concept of CAD Models
puting resources, millions of labeled data, and advanced
artificial intelligence methods, convolutional neural network This section briefly describes the basic concepts of
(CNN) has revolutionized image representation and computer-aided diagnosis (CAD) and transfer learning.
benefited a broad range of applications [18], including but Specifically, Figure 1 shows the flow chart of machine
not limited to object recognition [19], visual understanding learning- (ML-) based CAD and major architectures of
[20], and numerical regression [21, 22]. Quite different from CNN-based CAD. It should be noted that for diagnosis, a
conventional MBCD techniques, CNN attempts to integrate CAD model assumes the suspicious lesion regions have been
the feature extraction and lesion classification into a su- accurately delineated and its purpose is to predict the ma-
pervised learning procedure. The input of the CNN archi- lignancy of the input lesions.
tectures is image patches of outlined lesion regions, and its
output corresponds to the predicted lesion malignancy and
2.1. Computer-Aided Diagnosis (CAD). A CAD model can be
intuitively, time and labor can be reduced in feature engi-
used to provide additional information and support the
neering. Meanwhile, CNN is pushing forward the technique
decision making on disease diagnosis and cancer staging. It
upgrading in the field of medical imaging [23], medical
is different from a computer-aided detection model which
physics [24, 25], medical image analysis [26–28] and ra-
aims to detect, localize, or segment suspicious regions.
diotherapy [29, 30]. The research toward developing effec-
However, it should be noticed that a computer-aided de-
tive and efficient CNN-based MCBD models is still ongoing.
tection model can be placed ahead of a diagnosis model for
To the best of our knowledge, three review papers have
comprehensive analysis from the detection and localization
been published regarding deep learning based breast cancer
to the diagnosis of suspicious regions.
diagnosis. One concerns lesion detection and malignancy
prediction using mammography, ultrasound, magnetic
resonance imaging and digital tomosynthesis [31]. One 2.1.1. ML-Based CAD. A ML-based CAD model consists of
focuses on mammography and histology image processing feature extraction and machine learning-based classification
and analysis [32]. Meanwhile, it attempts to map the as shown in the left of Figure 1, and feature selection is
features/phenotypes between mammographic abnormalities optional. Widely used features come from image descriptors
and histological representation. The last one overviews deep that quantify the intensity, shape, and texture of a suspicious
learning in the detection and diagnosis of various kinds of region [34]. Preferred machine learning classifiers are not
cancers by using different imaging modalities [33]. In limited to artificial neural network (ANN), support vector
general, technical details in these review papers are not well machine (SVM), k-nearest neighbors, naive Bayesian, and
delivered. random forest (RF) [35]. Due to the emergency of radiomics
Input Input
Feature extraction Convolutional layers
CNN-based model
ML-based model
Feature selection Full connection layers
Machine learning-
based classification Other architectures
Output Output
(a) (b)
Figure 1: The diagram of the main flow chart of ML-based CAD (a) and major architectures of CNN-based CAD (b). The black dashed line
indicates the blocks are modifiable. The green dashed line denotes each step in the ML-based model is interpretable, and the red solid line
indicates the CNN-based model is data-driven when the architecture is fixed.
[36–38], it should be noted that feature selection becomes activation function in the corresponding layer i. Further-
more and more important and it aims to retrieve intrinsic more, how to design the architecture of deep learning
features of suspicious lesions. models in addition to the comprehensive analysis and
Mathematically, the procedure of using a pretrained ML- systematic methodologies of learning representation can be
based CAD model to predict the malignancy of a lesion can referred from [18, 19, 48].
be described as follows. First, an outlined suspicious region It should be noted that CNN models are data-driven and
(Ix) as the input is quantified with scalar variables (E(Ix )) by can be trained end-to-end. The models enable the in-
using feature extraction (E). Then, feature selection (S) is tegration of feature extraction, feature selection, and ma-
employed to decrease the feature dimension and to retrieve lignancy prediction into an optimization procedure.
informative features (S(E(Ix ))). In the end, the output of the Therefore, these retrieved features are not designed by
label (y) of the lesion (Ix) is predicted using machine learning human engineers but learned from the input data [19]. In
classifiers can be formulated as y � F(S(E(Ix ))). For general, remarkable performance of CNN-based CAD
comprehensive understanding, overviews regarding ma- models comes from advanced computing hardware resource
chine learning and breast cancer diagnosis can be referred to (i.e., graphic processing units and distributed computing
[8, 9]. system), open-source software, such as TensorFlow (https://
www.tensorflow.org/), and open challenges based on mil-
lions of high-quality labeled images, such as ImageNet
2.1.2. CNN-Based CAD. CNN models are computational (http://www.image-net.org/). Its success also benefits from
models that are composed of multiple processing layers to the novel design of architectures for deep learning, such as
retrieve features from raw data with multilevel representa- inception [43] and identity mapping [44].
tions and hierarchical abstraction [19]. As shown in the right
of Figure 1, a general architecture of CNN models is made up
of convolutional layers, full-connection layers, and pooling 2.2. Transfer Learning. Transfer learning, or knowledge
layers in addition to the input and output layers. Specifically, transfer, is more a machine learning strategy. It aims to reuse
Figure 2 shows the architecture of VGG16 which consists of a model pretrained in the source domain as a starting point
13 convolutional layers, 3 full-connection layers, 5 pooling in a different but related target domain [49]. In the field of
layers, and 1 softmax layer [39]. For further improvement in machine learning, an algorithm is typically designed to
object classification, many techniques can be embedded, address one isolated task, while through transfer learning,
including nonlinear filtering, data augmentation, local re- the algorithm can be further adapted to a new task (Figure 3).
sponse normalization, hyperparameter optimization, and It has several benefits using knowledge transfer. Above all,
multiscale representation [31, 32]. At present, widely used knowledge transfer enables the quality of the starting point
deep learning models include, but are not limited to, VGG in the target domain and thereby, promising results can be
[39], LeNet [40], AlexNet [41], GoogLeNet [42, 43], ResNet expected. Moreover, how to make use of a pretrained model
[44], YOLO [45], faster R-CNN [46] and LSTM [47]. is flexible. The model can be employed as a feature extractor
Mathematically, the procedure of using a pretrained for high-level representation of images and its parameters
CNN-based CAD model for the prediction of lesion ma- can be fine-tuned with target data. In addition, both time
lignancy can be described as following. Given a suspicious and cost can be reduced dramatically. Depending on com-
region (Ix), the output of a CNN-based model can be for- puting resources, it takes about days to months training a
malized as y � F(Ix ) � fn (fn−1 (· · · (f(Ix )))) where n deep model, while the time drops to hours when transferring
stands for the number of hidden layers and fi denotes the this model for target applications. Thanks to the accessibility
Input/output layer Full connection layer

Convolutional layer Softmax layer
Pooling layer
Figure 2: The architecture of VGG16. It consists of 13 convolutional layers, 3 full-connection layers, 5 pooling layers, and 1 softmax layer in
addition to the input and output layers.
of pretrained deep models online, high-cost hardware seems DREAM challenge is consisted of 82,000 images. More-
unnecessary. Most importantly, transfer learning relieves the over, among the public databases, BCDR-F03 is the only
requirement of huge amount of instances for model training, one consisted of FM images, while among in-house col-
which is critically helpful in medical imaging field. At present, lections, [55] is the one study that makes use of FM images
the most popular object classification is based on the (1655 FM images and 799 DM images), and all other
ImageNet [50] and without additional comments, pretrained databases and in-house collections are made up of DM
CNN models are all denoted an initialization on the ImageNet images.
in this study. Three public databases of DDSM (http://marathon.
csee.usf.edu/Mammography/Database.html), BCDR-F03
3. CNN-Based MBCD (http://bcdr.inegi.up.pt), INbreast (http://medicalresearch.
inescporto.pt/breastresearch/index.php), and MIAS (http://
This section firstly introduces the search strategy of litera- peipa.essex.ac.uk/info/mias.html) are accessible online, while
tures, involved databases and performance metrics. In the the DREAM challenge (https://www.synapse.org/#Synapse:
end, CNN-based MBCD methods are categorized into three syn4224222) is devoted to online competition and aims at
groups based on the design and use of CNN models. This improving the predictive accuracy of mammographic images
overview concentrates on peer-reviewed journal publica- for early detection and diagnosis of breast cancer. The IRMA
tions, and it provides technical details and pros and cons of [69] contains image patches selected from the DDSM, MIAS,
CNN models. and other two data sets. Among the public databases, DDSM
(“Digital Database for Screening Mammography”) remains
the largest available resource for mammographic image
3.1. Search Strategy for Literature Review. For the literature
analysis [70]. It consists of 14 volumes of benign lesion cases
survey, IEEEXplore, Pubmed, ScienceDirect, and Google
and 15 volumes of malignant lesion cases in addition to 2
Scholar were used to search publications relating to CNN-
volumes of benign lesion cases without callback. It also
based MBCD. The last update was at December 20, 2018.
contains 12 volumes of normal cases. The images in DDSM
Keywords are “convolutional neural network,” “deep learn-
are in an outdated image format with a bit depth of 12 or
ing,” “mammography,” “breast cancer,” and “diagnosis”.
16 bits per pixel, and image resolution is larger than [4000,
Specifically, only papers published on peer-reviewed journals
3000], both depending on scanners.
were selected, and our search yielded 18 research articles.
The database BCDR-F03 (“Film Mammography Dataset
Table 1 summarizes the literature from the used databases, the
Number 3”) is a subset of Breast Cancer Digital Repository
number (no.) of medical images in lesion classification and
(BCDR) that collects patient cases from the northern region
the diagnosis performance (AUC, the area under the curve;
of Portugal. It was made available for the development and
ACC, accuracy; SEN, sensitivity; SPE, specificity). Note that in
comparison of algorithms [52]. The BCDR-F03 contains 344
each literature, only the model which achieves the best
patient cases, 736 FM images, and 406 breast lesions. Among
classification performance is reported.
the lesions, 230 are benign (426 images) and 176 malignant
(310 images). Notably, BCDR-F03 contains FM images in
3.2. Involved Databases. Table 1 indicates that mostly used the gray-level digitized TIFF (Tagged Image File Format)
mammography databases come from in-house collection with a bit depth of 8 bits per pixel, and image resolution is
(7/18), followed by public databases of BCDR-F03 (5/18), [720, 1168].
DDSM (4/18), INbreast (3/18), MIAS (1/18), and IRMA The database INbreast is made up of 115 breast lesion
(1/18), and the last one comes from the DREAM challenge cases and 410 digital images [71]. However, only 56 cases are
(1/18). The number of medical images in databases ranges histologically verified (11 benign and 45 malignant lesions).
mainly from several hundreds to thousands. Notably, the The mammographic images are saved in DICOM (Digital
Target labels
Source labels
Target model
Source model Transfer learning
Target data
Source data
Figure 3: The diagram of knowledge transferred from the source domain to a different but related target domain. In the source domain, a
model is trained with sufficient high-quality instances (data and labels) and transfer learning enables the model used in a related target
domain. It relieves the requirement of huge amount of instances for the training of deep models in the target domain which is critically
helpful in medical imaging field.
Table 1: A summary of CNN based MBCD methods.

Year Database No. of images AUC ACC SPE SEN
[51] 2016 DDSM 600 0.967
[52] 2016 BCDR-F03 736 0.82 ± 0.03
[53] 2016 In-house 607 0.86
[54] 2017 In-house 3158 0.88 0.82 0.72 0.81
[55] 2017 In-house 2454 0.82 ± 0.02
[56] 2017 In-house 245 0.86 ± 0.01
[57] 2017 INbreast 115 0.91 ± 0.12 0.95 ± 0.05
[58] 2017 In-house 560 0.79 ± 0.02
[59] 2017 IRMA 2796 0.839 0.837 0.854 0.797
[60] 2018 In-house 78 0.81 ± 0.05
[61] 2018 In-house 3290 0.7274
[62] 2018 DDSM 600 0.974
MIAS 120 0.967
[63] 2018 BCDR-F03 736 0.813
[64] 2018 DDSM 5316 0.98 0.9735
BCDR-F03 600 0.96 0.9667
INbreast 200 0.97 0.9550
[65] 2018 DDSM 600 0.97
[66] 2018 DREAM 82,000 0.85
INbreast 115 0.95
[67] 2018 BCDR-F03 736 0.891 0.852
[68] 2018 BCDR-F03 736 0.88 0.81
Imaging and Communications in Medicine) format with 14- The MIAS database (“Mammographic Image Analysis
bit contrast resolution. The image matrix is [2560, 2238] or Society”) contains 322 digital images among which 67 le-
[3328, 4084] depending on imaging scanners. sions are benign and 53 lesions are malignant [72]. Quite
different from above-mentioned databases, MIAS provides the last full-connection layer of the frozen AlexNet. The
the image coordinates of center of each abnormality and the parameters in the AlexNet are initialized on the ImageNet
approximate radius (in pixels) of a circle to enclose the and keep unchanged, while the whole model is trained on
abnormality, but not the coordinates of points localized on medical instances. Reference [58] proposes a four-layered
the boundary of lesions. Images are stored 8 bits per pixel in model (3 convolutional layers and 1 full-connection layer)
the PGM (Probabilistic graphical model) format. The da- and a 4-fold cross-validation strategy is performed on 560
tabase has been reduced to a 200 micron pixel edge and lesions (280 benign and 280 malignant). Reference [62]
padded/clipped so that the image matrix is [1024, 1024]. designs a CNN architecture (5 convolutional layers and 2
full-connection layers), while it pretrains the model on the
ImageNet. Notably, parasitic metric learning is embedded
3.3. Performance Metrics. To quantify the classification that makes the best use of misclassified medical instances
performance of CAD models, widely used metrics are AUC and improves the diagnosis performance. Reference [65]
and ACC, followed by SEN and SPE (Table 1). Specifically, employs YOLO for lesion detection and localization fol-
ACC, SEN, and SPE are computed based on the confusion lowed by a tensor structure for the malignancy prediction.
matrix. As shown in Table 2, TP is the case which is his- And consequently, automatic detection and classification of
tologically verified positive and correctly predicted as suspicious lesions is achieved simultaneously. Similarly, [66]
“positive”, while FN represents the case histologically ver- uses the faster R-CNN for lesion detection and localization
ified positive but misclassified as “negative”. Furthermore, and the VGG for cancer diagnosis. The model is first trained
TN is the true negative case predicted correctly, and FP is the on the DDSM and further validated on the INbreast and the
true negative case but predicted as “positive” [73]. Generally, DREAM challenge. It performs as one of the best approaches
benign lesions are labeled with “negative” and malignant in mammographic image analysis. Reference [67] develops a
lesions are labeled “positive”. hybrid model. It first uses the pretrained GoogLeNet for
Given the labels and corresponding prediction results, feature extraction, and 3072 features are obtained. And then,
ACC, SEN, and SPE can be, respectively, formulated as an attention mechanism is proposed for feature selection. At
(TP + TN)/(TP + FN + FP + TN), TP/(TP + FN), and TN/ last, it uses LSTM to integrate both contexture information
(TN + FP). As to AUC, it is quantified based on the receiver from multiview image features and information of clinical
operating characteristics (ROC) curve. ROC is a curve of data for the lesion classification.
probability and AUC presents a model’s capacity of lesion Figure 4 demonstrates the flow chart and an example of
differentiation. To these 4 performance metrics, higher dedicated MBCD models. The flow chart highlights that the
values indicate better performance. CNN is a newly designed or modified network and the
example describes the architecture of the CNN model from
3.4. CNN-Based MBCD Models. In general, CNN-based [58]. It should be noted that parameters of dedicated models
models can be divided into dedicated models and trans- are with random initialization followed by iterative opti-
ferred models. The former include the proposal of new mization with medical instances.
architectures, the modification or integration of existing Although [55, 62, 66, 67] make use of the ImageNet for
CNN models, while the latter make the most use of pre- parameter initialization, it should be highlighted that one
trained models and further fine-tune them by using medical develops a new architecture [62], one modifies the existing
instances. Furthermore, it is found that some models just use architecture and introduces a new learning strategy [55], and
CNN for feature extraction and lesion diagnosis is fulfilled the others emphasize on the integration of two kinds of
by using machine learning classifiers. In particular, hand- network architectures for simultaneous detection and lo-
crafted features are taken into consideration. Therefore, in calization and final lesion diagnosis [66, 67]. Therefore,
this study, CNN-based MBCD models are broadly catego- [55, 62, 66, 67] are categorized into the group of dedicated
rized into three groups of dedicated models, transferred models.
models, and hybrid models. Table 3 summarizes the CNN-
based models from the model building to its pros and cons
analysis. Note that the pros of “parameter initialization” 3.4.2. Transferred CNN Models. Due to insufficient medical
indicate the model is pretrained with ImageNet. instances, deep CNN models pretrained on a large-scale of
labeled natural images (such as ImageNet) are transferred
and also fine-tuned with medical instances before the ap-
3.4.1. Dedicated MBCD Models. To enhance the diagnosis plication in breast cancer diagnosis. Reference [61] gives out
with unlabeled data, [54] proposes a graph-based semi- a systematic comparison of one shallow network (3 con-
supervised learning scheme, which is consisted of iterative volutional layers and 2 full-connection layers) and the
data weighting, feature selection, and data labeling before AlexNet. Transfer learning is concerned, and experiment
using the modified LeNet for lesion diagnosis. Experimental results indicate that CNN models with transfer learning
results indicate that the scheme requires quite a small outperform the models without transfer learning. Reference
portion of labeled data (100 lesions) for training and achieves [63] investigates three kinds of implementation of an 8-
promising performance on the unlabeled data (3058 lesions). layered CNN architecture. Parameters, such as the number
In addition, the scheme seems less sensitive to the initial of convolutional filters in each layer, are fine-tuned with
labeled data. Reference [55] adds 2 fully connected layers at mammographic lesion instances. Experimental comparison
Table 2: Confusion matrix.

Predicted positive Predicted negative
Histologically verified positive True positive (TP) False negative (FN)
Histologically verified negative False positive (FP) True negative (TN)
Table 3: Summary of CNN-based MBCD models from the model building to its pros and cons analysis.
Publication (year) Approach Pros (+)/cons (−)
(1) An 8-layered CNN +parameter initialization
[51] (2016) (2) SVM-based decision mechanism +decision mechanism
(3) Compared to ML- and CNN-based models −256 mid- and 2048 high-level features
(1) A 3-layered CNN +medical instances for training
[52] (2016) (2) SVM-based classification −17 low- and 400 high-level features
(3) Compared to ML- and CNN-based models −a shallow CNN model
(1) Transferred AlexNet +parameter initialization
[53] (2016) (2) SVM-based classification +soft-voting-based decision mechanism
(3) Classifier-based soft voting −29 low- and 3795 high-level features
(1) Modified LeNet +semisupervised learning
(2) Graph based semisupervised learning +a few labeled data used for training
[54] (2017)
(3) Feature dimension reduction +less sensitive to initial labeled data
(4) Using unlabeled data
(1) Modified AlexNet +parameter initialization
[55] (2017)
(2) Multitask transfer learning +improved generalizability
(1) Transferred the VGG +parameter initialization
[56] (2017) (2) SVM-based classification +decision mechanism
(3) Compared to ML- and CNN-based models −38 low- and 1472 high-level features
(1) R-CNN for detection and diagnosis +minimal user intervention in image analysis
[57] (2017) (2) Feature regression −781 low-level features for CNN feature regression
(3) RF-based classification
[58] (2017)
−a shallow CNN model
[59] (2017) (2) SVM-based classification +image analysis in transformed domain
(3) Data augmentation −a shallow CNN model
(1) VGG for feature extraction +2 features selected for diagnosis
[60] (2018) (2) Stepwise feature selection
(3) SVM-based classification
(1) Transferred AlexNet +parameter initialization
[61] (2018) (2) Data augmentation
(3) Compared to CNN models
(1) A 7-layered CNN +parameter initialization
[62] (2018)
(2) Parasitic metric learning +parasitic metric learning
(1) Transferred VGG +parameter initialization
[63] (2018)
(2) Compared to CNN-based models
(1) Transferred VGG/ResNet/Inception +parameter initialization
[64] (2018) (2) Comparison on 3 databases +systematic comparison
−time consuming
(1) YOLO and tensor structure +medical instances for training
[65] (2018)
(2) Data augmentation +simultaneous detection and classification
(1) Faster R-CNN and VGG +medical instances for training
[66] (2018) (2) Pretrained with the DDSM +both detection and diagnosis
+evaluated on a large-scale screening dataset
(1) GoogLeNet for feature extraction +medical instances for training
[67] (2018) (2) Attention mechanism for feature selection +multiview and clinical information fusion
(3) LSTM for feature fusion
(1) Transferred AlexNet/GoogLeNet +parameter initialization
[68] (2018) (2) Data augmentation
(3) Compared to ML- and CNN-based models
New or modified CNN

Flow chart Medical images Malignancy prediction
architecture
Dedicated MBCD models
An example

Pooling layer
Figure 4: The flow chart and an example of dedicated MBCD models. The flow chart highlights the CNN is a newly designed or modified
network, and the example describes the architecture of a CNN model in [58]. It should be noted that parameters of dedicated models are
with random initialization followed by iterative optimization with medical instances.
further indicates that incorporating handcrafted features decision mechanism is provided. After that, the MBCD
increases the classification performance. Reference [64] model integrates 256 midlevel and 2048 high-level features
concentrates the study on three deep learning models (VGG, for lesion classification. Reference [52] designs two shallow
RestNet, and GoogLeNet) and knowledge transfer is ex- networks and experimental results indicate the 3-layered
plored. Experiments are conducted to compare the random network (2 convolutional layers and 1 full-connection layer)
initialization and parameter initialization and to figure out obtains better performance. While for higher accuracy, SVM
how to fine-tune the models. Notably, three public databases is further employed which takes these CNN features as its
(DDSM, INbreast and MIAS) are analyzed. Reference [68] input. Experiment results show the diagnosis performance
compares two deep networks (AlexNet and GoogLeNet) achieves slight but significant improvement when 17 low-
which are pretrained on the ImageNet, two shallow CNN level and 400 high-level features are pooled for lesion
models, and two ML-based MBCD models. Experimental quantification. Reference [53] takes advantage of the pre-
results suggest that knowledge transfer is helpful in breast trained AlexNet for the lesion differentiation. More spe-
lesion diagnosis. cially, one SVM-based model uses 3795 high-level features as
Figure 5 shows the flow chart and an example of trans- its input and the other SVM-based model uses 29 low-level
ferred MBCD models. The flow chart highlights the offline features for the lesion classification. The outputs are fused by
training of a CNN model on nonmedical images, and soft voting and significant improvements are achieved in
moreover, it emphasizes fine-tuning the pretrained model with malignancy prediction. Reference [56] investigates different
medical instances. A representative example using VGG as the methodologies for feature fusion. It concerns 38 handcrafted
diagnosis model comes from [64]. It should be noted that features and 1472 CNN learned features, and SVM is as the
parameters of CNN architectures are predetermined in the classifier for each kind of feature. Then, the results from each
task of object recognition, and their values are further opti- SVM are fused for final decision making. The results show
mized toward mammographic breast lesion differentiation. that the integration of low- and high-level features signifi-
Existing deep architectures are made the most use of, and cantly improves cancer diagnosis. Reference [57] proposes a
these models [61, 63, 64, 68] are pretrained on the ImageNet hybrid framework for mammographic image analysis. With
and parameters are initialized. And then, mammographic minimal user intervention, it is capable of mass detection,
lesion instances are used to fine-tune the deep models. While lesion segmentation, and malignancy prediction. Specifi-
to further improve the diagnosis performance, additional cally, for lesion differentiation, it regresses the output of the
techniques, such as data augmentation, are embedded in the CNN model to 781 handcrafted features and then, a full-
training procedure. It should be noted that [61] has designed connection layer is added for feature abstraction. Finally, RF
shallow networks, while its purpose is to verify whether is utilized to improve the diagnosis accuracy. Reference [59]
transfer learning could improve the cancer diagnosis, and introduces a shallow network (2 convolutional layers and 1
thereby it is grouped into the transferred CNN models. full-connection layer). It alternatively cooperates with dis-
crete wavelet transform and curvelet transform for image
preprocessing. At last, a total of 784 features are handcrafted.
3.4.3. CNN Models as Feature Extractors. Among the CNN- Moreover, both softmax and SVM are compared, and SVM
based MBCD models, 7 out of 18 take CNN to retrieve high- outperforms softmax with slight increase. Reference [60]
level features for lesion representation. Reference [51] de- takes advantage of 1472 high-level features from the pre-
velops an 8-layered network (5 convolutional layers and 3 trained VGG with frozen parameters. Its novelty comes from
full-connection layers). The model is pretrained on the the proposal of step-wise feature selection and the 2 most
ImageNet to overcome the issue of limited medical in- frequently selected features are used for SVM-based breast
stances. And then, SVM performs as the classifier and a lesion classification.
Nature images Object classification
Flow chart CNN architecture
Medical images Fine tuning Malignancy prediction

Transferred CNN models
An example

Pooling layer Fine tuning
Figure 5: The flow chart and an example of transferred MBCD models. The flow chart emphasizes transfer learning (dashed arrows) and
fine-tuning, and the example comes from [64] which makes use of pretrained VGG16 for malignancy prediction. It should be noted that
parameters of pretrained models are well-determined in the source domain, while fine-tuning attempts to use medical instances for further
optimization of these parameters toward the target task.
Figure 6 shows the flow chart and an example of CNN “mammography” images. The models are generally divided
models as feature extractors. The flow chart highlights in- into three groups (Table 4): one highlights the design of new
formation fusion. In other words, whether a CNN model is architectures or the modification or integration of existing
newly designed or pretrained becomes not important and networks (Figure 4); one concentrates on the use of transfer
using low-level feature is optional. Information fusion can be learning and fine-tuning in breast cancer diagnosis (Figure 5);
divided into two approaches. One is feature fusion followed and the last one concerns a hybrid model in which CNN
by a classifier, and the other is decision fusion of lesion performs for feature extraction and information fusion be-
malignancy predicted by using one or more classifiers. The comes indispensable in decision making (Figure 6). In ad-
example comes from [51] which develops a new CNN model dition, Table 3 summarizes these models from the model
and the model is pretrained on ImageNet. At last, the model building to its pros and cons analysis.
fuses the prediction results (decision fusion) from SVM
classifiers which separately use 384 midlevel features and
2048 high-level features as the input. 4.1. Our Findings. To overcome the issue of limited medical
Prior studies have proved the benefits of low-level fea- instances, there are 10 publications that employ transfer
tures in mammographic image analysis. And at present, how learning [51, 53, 55, 56, 61–64, 66, 68], with or without fine-
to select the informative CNN features [60] and how to fuse tuning. Transfer learning is able to alleviate this issue to some
low-, mid-, and high-level features and clinical information extent, since deep models have been optimized using
have become important [52, 53, 56]. It should be mentioned massive amount of data in the source domain; and conse-
that even if some MBCD models concern handcrafted quently, the time and labor can be considerably reduced in
features [53, 56], the ultimate purpose is to construct a the target domain. In particular, it has been verified that
hybrid framework for improved diagnosis and thereby, these transfer learning benefits the differentiation of breast lesions
publications [53, 56] are categorized into the third group. seen in mammographic images. Besides, to increase the
number of medical instances, data augmentation is used
[59, 61, 65, 68]. It makes sense in lesion malignancy pre-
3.4.4. Technical Highlights among CNN-Based MBCD diction, since a lesion might be presented in any particular
Models. Table 4 summarizes the technical highlights that orientation in screening and thus, the MBCD model should
can distinguish each kind of CNN-based MBCD models. In be able to learn and recognize the lesion malignancy. For
the Table, “✔” indicates the distinct component in the data augmentation, besides image rotation and flipping,
model, “✖” denotes the component is not included in the other techniques can be adapted, such as image quality
models, while “—” means the component is not important in degrading (https://github.com/aleju/imgaug) and image
this kind of CNN-based models. deformation [74–76].
To improve the diagnosis performance, 11 out of 18
4. Discussion publications develop shallow architectures or modify
existing networks [51, 52, 54, 57–60, 62, 65–67]. Shallow
A total of 18 peer-reviewed journal publications (Table 1) are architectures decrease the number of medical instances for
found with regard to the “convolutional neural network” or training, while machine learning classifiers should be uti-
“deep learning” based “breast cancer diagnosis” using lized when modified deep networks with frozen or fine-
Information fusion
Feature engineering Low-level features
Feature fusion A classifier Malignancy
Flow chart Medical images prediction
CNN models as feature extractors
Midlevel and/or
CNN architecture high-level features Classifiers Decision fusion
384 midlevel features
2048 high-level features
An example

Pooling layer Information fusion
Figure 6: The flow chart and an example of CNN performing as feature extractors. The flow chart highlights the information fusion which
can be further divided into two approaches, feature fusion followed by a classifier or decision fusion of lesion malignancy predicted by using
one or more classifiers. The example comes from [51] which develops a new CNN model and the model is pretrained on ImageNet. At last,
the model fuses the prediction results from SVM classifiers which separately use 384 midlevel features and 2014 high-level features as its
input.
Table 4: Technical highlights.

New architecture Transfer learning Fine-tuning Information fusion
Dedicated CNN models ✔ ✖ ✖ ✖
Transferred CNN models — ✔ ✔ ✖
CNN models as feature extractors — — — ✔
tuned parameters perform as feature extractors. However, provide the comparison between CNN- and ML-based
problems occur. The first problem concerns which classifier models [51, 52, 56, 68] and ML-based models are treated
to be applied for the differentiation of benign and malignant as the baseline. It should be noted that ML-based models
lesions. It is found that 9 out of 11 publications select SVM benefit from the prior knowledge and clinical experience in
[51, 52, 54, 58–60, 62, 65, 66], and 1 uses RF [57] and 1 feature crafting, feature selection, and the use of machine
chooses LSTM [67] for malignancy prediction. The second learning classifiers. In particular, it is feasible to build a ML-
one is how to choose informative and predictive features based model on a very small database [36]. Besides, ML-
among hundreds to thousands of variables. Most publica- based models are relatively lightweight computing and re-
tions address this question by comprehensive experiments to quire no specific hardware and thus, these models can be
make a trade-off between the diagnosis efficiency and ef- easily deployed and managed in daily work.
fectiveness, while only [56] proposes using the frequency of Integrating multiple representation of mammographic
the CNN feature selected in the training stage as the lesions can enhance the performance of breast cancer di-
weighting of the feature importance. Last but not the least, it agnosis, while how to incorporate low-, mid-, and high-level
is time-consuming and troublesome. In general, it takes days features or multiview data is quite difficult. There are 4
to weeks to develop new architectures and to modify or to publications [51, 53, 56, 67] which provide mechanism for
integrate deep models due to the requirements of model information fusion or decision fusion. Reference [51] pro-
training, parameter optimization, feature selection, and al- poses a decision mechanism by evaluating the consistency of
gorithm comparison. the results from the midlevel features and the high-level
It is also found that 7 publications consider low-level features. If not consistent, gray information would be added
and/or clinical features [51–54, 56, 59, 67]. Low-level fea- to assess the similarity and support the decision making.
tures are mainly derived from intensity statistics, shape Both [53, 56] build ensemble classifiers by averaging the
description, and texture analysis [34]. Specifically, these results from two SVM classifiers among which one makes
features can be further analyzed with multiscale de- use the pretrained CNN features and the other analyzes
composition or in transform spaces. Clinical information handcrafted features. Reference [67] utilizes LSTM cells to
includes breast density, patients’ age, and other symptoms, integrate the features from multiview data. Since multiview
such as microcalcification. In addition, 4 publications data contain contextual information, the variations among
multiview images may contribute additional information in acquisition further degrade the imaging quality. Therefore,
lesion interpretation. quantitative image quality assessment may be necessary [86].
Moreover, due to different shapes and margin of suspicious
lesions and also ambiguous boundaries between lesions and
4.2. Technical Challenges. Several technical challenges re- surrounding tissues, the quality of lesion delineation is
main. The first challenge comes from how to use the pre- unstable, and thereby, the techniques for automatic mam-
trained deep CNN models which is closely related to the mographic breast lesion detection and segmentation are still
MBCD performance [77, 78]. However, there is no definitive in need of improvement [87]. In addition, evolutionary
answer on how to fine-tune the network and how many pruning of knowledge transfer of deep models that are
medical instances is sufficient for the fine-tuning, even good pretrained on sufficient medical images is promising for
practice is available [79]. The simplest way is to take the mammographic breast lesion diagnosis because of the
parameters of the whole network or some layers of the similar feature space [88]. Last but not the least, it is always
network tunable. Some studies suggest layer-wise fine- desirable to build a seamless system to localize the suspicious
tuning, while the time consumption will be dramatically lesions and give out the malignancy prediction simulta-
increased [80]. On the other hand, when using deep models neously [65, 66].
as feature extractors, other technical issues arise, including
how to select high-level features, how to integrate multi-
perspective information, and which machine learning 4.3. Future Focus. Except for the technical challenges
classifier is employed. It is pitiful that no tutorial or practical aforementioned, another three topics should be focused on
guidelines are repeatable. In clinic, to improve the perfor- in the future work. The first one is to collect sufficient high-
mance of breast cancer diagnosis, various imaging modal- quality mammographic instances. Due to the limited
ities and clinical data are taken into account that further funding, scarce medical expertise, and privacy issues, there is
imposes difficulties on information fusion [9]. Since no one- no big leap in data sharing, in particular, the mammographic
size-fits-all solution is available, prior knowledge, previous lesion images. At present, the DDSM remains the largest
studies, and empirical experience become more and more publicly available database as well as the first choice in large-
important to address these technical issues [78–83]. scale mammographic image analysis [89]. While based
It is also challenging on how to avoid overfitting in the on the fact that over 150 million mammographic exami-
optimization of deep networks. Dropout is proposed to nations are performed worldwide per year, there is signif-
address the problem [84] which aims to randomly drop units icant room for improvement in data collection and sharing.
(along with connections) from the network in the training In particular, lack of imaging data restricts the develop-
stage. It can prevent units from coadapting too much, and a ment and upgrading of intelligent systems for personal-
practical guide is provided for the training of a Dropout ized diagnosis, including but not limited to the design of
network [84]. It is full of potential to avoid overfitting by deeper architectures, hyperparameter optimization, and the
increasing the number of medical instances for training. At evaluation of generalization capacity. Fortunately, rapid
last, if there is no possibility to reduce the architecture progress is seen in the era of big data and many public
complexity and no way to increase the number of training databases have been released online, such as TCIA (http://
instances, the mainstream is to manipulate parameters, such www.cancerimagingarchive.net/), and various challenges are
as the learning rate, and to monitor the drop of performance open, such as the DREAM challenge. With such a stan-
metrics between the training phase and the validation phase dardization, it will become easier to compare different ap-
[58, 60, 61, 68]. It also should be mentioned that the proaches on the same problem of the same database and
threshold of the drop is subjective, and thus, comprehensive thereby pushing forward the techniques of CNN-based
experiments become necessary. MBCD.
The third challenge is the curse of dimensionality [85]. It Another topic is about the interpretation of the learned
is known that the primary purpose of deep learning is for CNN features. In contrast to handcrafted features with
recognizing the target from thousands of object categories. mathematical formalization and clear explanation, the in-
However, MBCD is a binary classification problem, and the terpretation of retrieved CNN features is quite poor. One
lesions seen in mammographic images are to be labeled as way to tackle this issue is from qualitative understanding
benign or malignant. Thus, it seems not convincing to use [55, 58] based on visualization. Reference [90] provides a
thousands of features for a binary classification problem technique for layer-wise feature visualization. In object
regarding hundreds of medical instances [51–53, 56]. Some recognition, the technique indicates that shallow layers
studies take recourse to feature selection [60] and feature typically represent the presence of edges, middle layers
dimension reduction [54]. As to deep networks, the fre- mainly detects motifs by spotting particular arrangements of
quency of features selected in the training phase as a fine structures, while deep layers attempt to assemble these
weighting factor of feature importance is meaningful [60]. motifs into a larger cluster to be a part of or the whole object
In practice, challenges exist in each step of the building [19, 58]. It should be admitted that the layer-wise visuali-
of CNN-based MBCD models. First, a number of factors zation technique facilitates the visual perception and further
influence the quality of mammographic imaging, such as the understanding of what the networks have learned. Reference
imaging scanner and reconstruction methods, and both [91] analyzed the predicted results in two-dimensional space
breast compression and motion artifacts during image using t-distributed stochastic neighbor embedding (t-SNE).
The t-SNE represents each object by a point in a scatter plot CNN-based MBCD models [103–105]. In addition, some
where nearby points denote similar objects and distant technical details, such as how to prepare medical instances
points indicate dissimilar objects. Therefore, a clear insight is for training, are not delivered in this review, while it should
provided into the underlying structure of malignancy pre- be kept in mind that each step is related to mammographic
diction [55]. Quantitative interpretation of deep learning is image analysis.
ongoing. Reference [92] gives a geometric view to un-
derstand the success of deep learning. They claim that the 5. Conclusion
fundamental principle attributing to the success is the
manifold structure in data, and deep learning can learn the This study presents a technical review of the recent progress
manifold and the probability distribution on it. Reference of CNN-based MBCD. It categorizes the techniques into
[93] provides theory on how to interpret the concept learned three groups based on how to use CNN models. Further-
and the decision made by a deep model. It further discusses a more, the findings from the model building to the pros and
number of questions in interpretability, technical challenges, cons of each model are summarized. In addition, technical
and possible applications. The third topic is the translation of challenges, future focus, and limitations are pointed out. At
the clinical research of CNN-based MBCD into the decision present, the design and use of CNN-based MBCD is at its
supporting in clinical practice. There is no doubt that deep early stage and result-oriented. To the ultimate goal of using
learning tools can provide valuable and accurate in- deep learning tools to facilitate clinical practice, there is still a
formation for cancer diagnosis, while it is impossible to take long way ahead. This review benefits scientific researcher,
the role and responsibility of clinicians. The fundamental industrial engineers, and those who are devoted to intelligent
role of a clinician in routine work is to collaborate with other cancer diagnosis.
team members, including physicians, technologists, nurses,
therapists, and even patients [94]. Thus, before accepting Conflicts of Interest
these decision-supporting systems for daily use, it should
provide profound understanding and visual interpretation The authors declare that they have no conflicts of interest
of deep learning tools, not only the surpassing human-level regarding the publication of this paper.
performance.
Furthermore, one big step to use CNN-based MBCD Authors’ Contributions
models for clinical applications comes from the review and
Lian Zou and Shaode Yu contributed equally to this work.
approval from the Food and Drug Administration (FDA). To
date, several FDA-approved CAD systems have been in the
market, such as the QVCAD system (QView Medical Inc, Acknowledgments
Los Altos, CA) that uses deep learning for automated 3D The authors would like to thank the editor and reviewers for
breast ultrasound analysis. With the increasing use of deep their constructive comments that have helped to improve
learning algorithms, more and more CNN-based CAD the paper quality. Thanks are also given to those researchers
systems will be approved by the FDA. Basically, compelling who share datasets and codes for fair comparison. This work
properties, such as expert-level performance, robustness, is supported in part by grants from the National Key Re-
and generalizability, should be guaranteed on different search and Develop Program of China (2016YFC0105102),
imaging devices. While from the perspective of long-term the Leading Talent of Special Support Project in Guangdong
evolution, a global real-life application accounting for (2016TX03R139), Fundamental Research Program of
widespread geographic, ethic, and genetic variations should Shenzhen (JCYJ20170413162458312), the Science Founda-
be considered. Therefore, there is still a long way ahead of the tion of Guangdong (2017B020229002, 2015B020233011,
translation of deep learning tools from scientific research to 2014A030312006) and the Beijing Center for Mathematics
clinical practice. and Information Interdisciplinary Sciences, and the Na-
tional Natural Science Foundation of China (61871374).
4.4. Limitations. There are several limitations. First, this References
review focuses on CNN for automated MBCD. For
computer-aided MBCD, it can also be well tackled by using [1] R. L. Siegel, K. D. Miller, and A. Jemal, “Cancer statistics,
other CAD techniques, such as case retrieval [95–97] and 2016,” CA: A Cancer Journal for Clinicians, vol. 66, no. 1,
breast density estimation [98, 99]. Moreover, this study pp. 7–30, 2016.
concerns only mammography. For comprehensive disease [2] L. Fan, K. Strasser-Weippl, J.-J. Li et al., “Breast cancer in
analysis, other imaging modalities, such as ultrasound and China,” Lancet Oncology, vol. 15, no. 7, pp. e279–e289, 2014.
magnetic resonance, should be taken into consideration [31]. [3] L. Tabár, A. Gad, L. H. Holmberg et al., “Reduction in
mortality from breast cancer after mass screening with
Besides, this review is limited to two-dimensional image
mammography,” The Lancet, vol. 325, no. 8433, pp. 829–832,
analysis, and many other medical tasks use CNN models to 1985.
tackle volumetric images [100–102]. In particular, this study [4] S. Yu, S. Wu, L. Zhuang et al., “Efficient segmentation of a
concerns only peer-reviewed journal publications that breast in B-mode ultrasound tomography using three-
considerably reduces the number of publications for analysis dimensional GrabCut (GC3D),” Sensors, vol. 17, no. 8,
and consequently, it might omit some high-quality p. 1827, 2017.
[5] R. Longo, F. Arfelli, R. Bellazzini et al., “Towards breast [20] Y. Guo, Y. Liu, A. Oerlemans, S. Lao, S. Wu, and M. S. Lew,
tomography with synchrotron radiation at Elettra: first “Deep learning for visual understanding: a review,” Neu-
images,” Physics in Medicine and Biology, vol. 61, no. 4, rocomputing, vol. 187, pp. 27–48, 2016.
pp. 1634–1649, 2016. [21] S. Yu, S. Wu, L. Wang, F. Jiang, Y. Xie, and L. Li, “A shallow
[6] M. J. Michell, A. Iqbal, R. K. Wasan et al., “A comparison of convolutional neural network for blind image sharpness
the accuracy of film-screen mammography, full-field digital assessment,” PloS one, vol. 12, no. 5, Article ID e0176632,
mammography, and digital breast tomosynthesis,” Clinical 2017.
Radiology, vol. 67, no. 10, pp. 976–981, 2012. [22] Y. Li, Z. Wang, G. Dai, S. Wu, S. Yu, and Y. Xie, “Evaluation
[7] F. M. Hall, J. M. Storella, D. Z. Silverstone, and G. Wyshak, of realistic blurring image quality by using a shallow con-
“Nonpalpable breast lesions: recommendations for biopsy volutional neural network,” in Proceedings of the 2017 IEEE
based on suspicion of carcinoma at mammography,” Ra- International Conference on Information and Automation
diology, vol. 167, no. 2, pp. 353–358, 1988. (ICIA), vol. 1, pp. 853–857, Macau, China, July 2017.
[8] J. Tang, R. M. Rangayyan, J. Xu, I. El Naqa, and Y. Yang, [23] Z. Zhang, X. Liang, X. Dong, Y. Xie, and G. Cao, “A sparse-
“Computer-aided detection and diagnosis of breast cancer view CT reconstruction method based on combination of
with mammography: recent advances,” IEEE Transactions on DenseNet and deconvolution,” IEEE Transactions on Med-
Information Technology in Biomedicine, vol. 13, no. 2, ical Imaging, vol. 37, no. 6, pp. 1407–1417, 2018.
pp. 236–251, 2009. [24] Y. Liu, S. Stojadinovic, B. Hrycushko et al., “A deep con-
[9] N. I. R. Yassin, S. Omran, E. M. F. El Houby, and H. Allam, volutional neural network-based automatic delineation
“Machine learning techniques for breast cancer computer strategy for multiple brain metastases stereotactic radio-
aided diagnosis using different image modalities: a sys- surgery,” PloS One, vol. 12, no. 10, Article ID e0185844, 2017.
tematic review,” Computer Methods and Programs in Bio- [25] R. Wang, X. Liang, X. Zhu, and Y. Xie, “A feasibility of
medicine, vol. 156, pp. 25–45, 2018. respiration prediction based on deep Bi-LSTM for real-time
[10] Y. Jiang, R. M. Nishikawa, R. A. Schmidt, C. E. Metz, tumor tracking,” IEEE Access, vol. 6, pp. 51262–51268, 2018.
M. L. Giger, and K. Doi, “Improving breast cancer diagnosis [26] D. Shen, G. Wu, and H.-I. Suk, “Deep learning in medical
with computer-aided diagnosis,” Academic radiology, vol. 6, image analysis,” Annual Review of Biomedical Engineering,
no. 1, pp. 22–33, 1999. vol. 19, no. 1, pp. 221–248, 2017.
[11] H. D. Cheng, J. Shan, W. Ju, Y. Guo, and L. Zhang, “Au- [27] W. Qin, J. Wu, F. Han et al., “Superpixel-based and
tomated breast cancer detection and classification using boundary-sensitive convolutional neural network for auto-
ultrasound images: a survey,” Pattern Recognition, vol. 43, mated liver segmentation,” Physics in Medicine & Biology,
no. 1, pp. 299–317, 2010. vol. 63, no. 9, article 095017, 2018.
[12] A. Jalalian, S. B. T. Mashohor, H. R. Mahmud, [28] T. Xiao, L. Liu, K. Li, W. Qin, S. Yu, and Z. Li, “Comparison
M. I. B. Saripan, A. R. B. Ramli, and B. Karasfi, “Computer- of transferred deep neural networks in ultrasonic breast
aided detection/diagnosis of breast cancer in mammography masses discrimination,” BioMed Research International,
and ultrasound: a review,” Clinical Imaging, vol. 37, no. 3, vol. 2018, Article ID 4605191, 9 pages, 2018.
pp. 420–426, 2013. [29] P. Meyer, V. Noblet, C. Mazzara, and A. Lallement, “Survey
[13] S. Mambou, P. Maresova, O. Krejcar, A. Selamat, and on deep learning for radiotherapy,” Computers in Biology
K. Kuca, “Breast cancer detection using infrared thermal and Medicine, vol. 98, pp. 126–146, 2018.
imaging and a deep learning model,” Sensors, vol. 18, no. 9, [30] X. Zhen, J. Chen, Z. Zhong et al., “Deep convolutional neural
p. 2799, 2018. network with transfer learning for rectum toxicity prediction
[14] H. D. Cheng, X. Cai, X. Chen, L. Hu, and X. Lou, “Computer- in cervical cancer radiotherapy: a feasibility study,” Physics in
aided detection and classification of microcalcifications in Medicine & Biology, vol. 62, no. 21, pp. 8246–8263, 2017.
mammograms: a survey,” Pattern Recognition, vol. 36, no. 12, [31] J. R. Burt, N. Torosdagli, N. Khosravan et al., “Deep learning
pp. 2967–2991, 2003. beyond cats and dogs: recent advances in diagnosing breast
[15] R. M. Rangayyan, F. J. Ayres, and J. E. Leo Desautels, “A cancer with deep neural networks,” British Journal of Ra-
review of computer-aided diagnosis of breast cancer: toward diology, vol. 91, article 20170545, 2018.
the detection of subtle signs,” Journal of the Franklin In- [32] A. Hamidinekoo, E. Denton, A. Rampun, K. Honnor, and
stitute, vol. 344, no. 3-4, pp. 312–348, 2007. R. Zwiggelaar, “Deep learning in mammography and breast
[16] I. Christoyianni, A. Koutras, E. Dermatas, and histology, an overview and future trends,” Medical Image
G. Kokkinakis, “Computer aided diagnosis of breast cancer Analysis, vol. 47, pp. 45–67, 2018.
in digitized mammograms,” Computerized Medical Imaging [33] Z. Hu, J. Tang, Z. Wang, K. Zhang, L. Zhang, and Q. Sun,
and Graphics, vol. 26, no. 5, pp. 309–319, 2002. “Deep learning for image-based cancer detection and
[17] L. Wei, Y. Yang, R. M. Nishikawa, and Y. Jiang, “A study on diagnosis−a survey,” Pattern Recognition, vol. 83, pp. 134–
several machine-learning methods for classification of ma- 149, 2018.
lignant and benign clustered microcalcifications,” IEEE [34] D. C. Moura and M. A. Guevara López, “An evaluation of
Transactions on Medical Imaging, vol. 24, no. 3, pp. 371–380, image descriptors combined with clinical data for breast
2005. cancer diagnosis,” International Journal of Computer
[18] Y. Bengio, A. Courville, and P. Vincent, “Representation Assisted Radiology and Surgery, vol. 8, no. 4, pp. 561–574,
learning: a review and new perspectives,” IEEE Transactions 2013.
on Pattern Analysis and Machine Intelligence, vol. 35, no. 8, [35] G. An, K. Omodaka, S. Tsuda et al., “Comparison of ma-
pp. 1798–1828, 2013. chine-learning classification models for glaucoma manage-
[19] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” ment,” Journal of Healthcare Engineering, vol. 2018, Article
Nature, vol. 521, no. 7553, pp. 436–444, 2015. ID 6874765, 8 pages, 2018.
[36] R. J. Gillies, P. E. Kinahan, and H. Hricak, “Radiomics: [53] B. Q. Huynh, H. Li, and M. L. Giger, “Digital mammographic
images are more than pictures, they are data,” Radiology, tumor classification using transfer learning from deep
vol. 278, no. 2, pp. 563–577, 2016. convolutional neural networks,” Journal of Medical Imaging,
[37] S. S. F. Yip and H. J. W. L. Aerts, “Applications and limi- vol. 3, no. 3, article 034501, 2016.
tations of radiomics,” Physics in Medicine and Biology, [54] W. Sun, T.-L. Tseng, J. Zhang, and W. Qian, “Enhancing
vol. 61, no. 13, pp. R150–R166, 2016. deep convolutional neural network scheme for breast cancer
[38] X. Xu, X. Zhang, Q. Tian et al., “Quantitative identification of diagnosis with unlabeled data,” Computerized Medical Im-
nonmuscle-invasive and muscle-invasive bladder carcino- aging and Graphics, vol. 57, pp. 4–9, 2017.
mas: a multiparametric MRI radiomics analysis,” Journal of [55] R. K. Samala, H.-P. Chan, L. M. Hadjiiski, M. A. Helvie,
Magnetic Resonance Imaging, vol. 278, no. 2, pp. 563–577, K. Cha, and C. Richter, “Multi-task transfer learning deep
2018. convolutional neural network: application to computer-
[39] K. Simonyan and A. Zisserman, “Very deep convolutional aided diagnosis of breast cancer on mammograms,” Phys-
networks for large-scale image recognition,” 2014, https:// ics in Medicine & Biology, vol. 62, no. 23, p. 8894, 2017.
arxiv.org/abs/1409.1556. [56] N. Antropova, B. Q. Huynh, and M. L. Giger, “A deep feature
[40] Y. LeCun, B. Boser, J. S. Denker et al., “Backpropagation fusion methodology for breast cancer diagnosis demon-
applied to handwritten zip code recognition,” Neural strated on three imaging modality datasets,” Medical Physics,
Computation, vol. 1, no. 4, pp. 541–551, 1989. vol. 44, no. 10, pp. 5162–5171, 2017.
[41] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet [57] N. Dhungel, G. Carneiro, and A. P. Bradley, “A deep learning
classification with deep convolutional neural networks,” in approach for the analysis of masses in mammograms with
Proceedings of the Advances in Neural Information Processing minimal user intervention,” Medical Image Analysis, vol. 37,
Systems, pp. 1097–1105, Lake Tahoe, NV, USA, December pp. 114–128, 2017.
2012. [58] Y. Qiu, S. Yan, R. R. Gundreddy et al., “A new approach to
[42] C. Szegedy, W. Liu, Y. Jia et al., “Going deeper with con- develop computer-aided diagnosis scheme of breast mass
volutions,” in Proceedings of the 2015 IEEE Conference on classification using deep learning technology,” Journal of
Computer Vision and Pattern Recognition (CVPR), pp. 1–9, X-Ray Science and Technology, vol. 25, no. 5, pp. 751–763, 2017.
Boston, MA, USA, June 2015. [59] M. M. Jadoon, Q. Zhang, I. Ul Haq, S. Butt, and A. Jadoon,
[43] C. Szegedy, V. Vanhoucke, S. Ioffe et al., “Rethinking the “Three-class mammogram classification based on descriptive
inception architecture for computer vision,” in Proceedings
CNN features,” BioMed Research International, vol. 2017,
of the IEEE Conference on Computer Vision and Pattern
Article ID 3640901, 11 pages, 2017.
Recognition, pp. 2818–2826, Las Vegas, NV, USA, June 2016.
[60] K. Mendel, H. Li, D. Sheth, and M. Giger, “Transfer learning
[44] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning
from convolutional neural networks for computer-aided
for image recognition,” in Proceedings of the 2016 IEEE
diagnosis: a comparison of digital breast tomosynthesis
Conference on Computer Vision and Pattern Recognition
and full-field digital mammography,” Academic Radiology,
(CVPR), pp. 770–778, Las Vegas Valley, NV, USA, June 2016.
2018, In press.
[45] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only
[61] X. Zhang, Y. Zhang, E. Y. Han et al., “Classification of whole
look once: Unified, real-time object detection,” in Pro-
mammogram and tomosynthesis images using deep con-
ceedings of the 2016 IEEE Conference on Computer Vision and
volutional neural networks,” IEEE Transactions on Nano-
Pattern Recognition (CVPR), pp. 779–788, Las Vegas Valley,
NV, USA, June 2016. bioscience, vol. 17, no. 3, pp. 237–242, 2018.
[46] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: to- [62] Z. Jiao, X. Gao, Y. Wang, and J. Li, “A parasitic metric
wards real-time object detection with region proposal net- learning net for breast mass classification based on mam-
works,” in Proceedings of the Advances in Neural Information mography,” Pattern Recognition, vol. 75, pp. 292–301, 2018.
Processing Systems, pp. 91–99, Montreal, Canada, December [63] A. C. Perre, L. A. Alexandre, and L. C. Freire, “Lesion
2015. classification in mammograms using convolutional neural
[47] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence networks and transfer learning,” Computer Methods in
learning with neural networks,” in Proceedings of the Ad- Biomechanics and Biomedical Engineering: Imaging & Vi-
vances in Neural Information Processing Systems, pp. 3104– sualization, pp. 1–7, 2018.
3112, Montreal, Canada, December 2014. [64] H. Chougrad, H. Zouaki, and O. Alheyane, “Deep con-
[48] J. Schmidhuber, “Deep learning in neural networks: an volutional neural networks for breast cancer screening,”
overview,” Neural Networks, vol. 61, pp. 85–117, 2015. Computer Methods and Programs in Biomedicine, vol. 157,
[49] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE pp. 19–30, 2018.
Transactions on Knowledge and Data Engineering, vol. 22, [65] M. A. Al-masni, M. A. Al-antari, J.-M. Park et al., “Simul-
no. 10, pp. 1345–1359, 2010. taneous detection and classification of breast masses in
[50] O. Russakovsky, J. Deng, H. Su et al., “ImageNet large scale digital mammograms via a deep learning YOLO-based CAD
visual recognition challenge,” International Journal of system,” Computer Methods and Programs in Biomedicine,
Computer Vision, vol. 115, no. 3, pp. 211–252, 2015. vol. 157, pp. 85–94, 2018.
[51] Z. Jiao, X. Gao, Y. Wang, and J. Li, “A deep feature based [66] D. Ribli, A. Horváth, Z. Unger, P. Pollner, and I. Csabai,
framework for breast masses classification,” Neuro- “Detecting and classifying lesions in mammograms with
computing, vol. 197, pp. 221–231, 2016. deep learning,” Scientific Reports, vol. 8, no. 1, pp. 85–94,
[52] J. Arevalo, F. A. González, R. Ramos-Pollán, J. L. Oliveira, 2018.
and M. A. Guevara Lopez, “Representation learning for [67] H. Wang, J. Feng, Z. Zhang et al., “Breast mass classification
mammography mass lesion classification with convolutional via deeply integrating the contextual information from
neural networks,” Computer Methods and Programs in multi-view data,” Pattern Recognition, vol. 80, pp. 42–52,
Biomedicine, vol. 127, pp. 248–257, 2016. 2018.
[68] S. Yu, L. L. Liu, Z. Y. Wang, G. Z. Dai, and Y. Q. Xie, [84] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
“Transferring deep neural networks for the differentiation of R. Salakhutdinov, “Dropout: a simple way to prevent neural
mammographic breast lesions,” China Technological Sci- networks from overfitting,” Journal of Machine Learning
ences, vol. 62, no. 3, pp. 441–447, 2018. Research, vol. 15, no. 1, pp. 1929–1958, 2014.
[69] J. E. Oliveira, M. O. Gueld, A. d. A. Araújo, B. Ott, and [85] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman,
T. M. Deserno, “Toward a standard reference database for “Return of the devil in the details: delving deep into con-
computer-aided mammography,” in Proceedings of the volutional nets,” 2014, https://arxiv.org/abs/1405.3531.
Medical Imaging 2008: Computer-Aided Diagnosis, vol. 6915, [86] Z. Zhang, G. Dai, X. Liang, S. Yu, L. Li, and Y. Xie, “Can
p. 69151Y, San Diego, CA, USA, March 2008. signal-to-noise ratio perform as a baseline indicator for
[70] M. Heath, K. Bowyer, D. Kopans et al., “The digital database medical image quality assessment,” IEEE Access, vol. 6,
for screening mammography,” in Proceedings of the In- pp. 11534–11543, 2018.
ternational Workshop on Digital Mammography, vol. 6915, [87] A. Oliver, J. Freixenet, J. Martı́ et al., “A review of automatic
pp. 212–218, Toronto, Canada, June 2000. mass detection and segmentation in mammographic im-
[71] I. C. Moreira, I. Amaral, I. Domingues, A. Cardoso, ages,” Medical Image Analysis, vol. 14, no. 2, pp. 87–110,
M. J. Cardoso, and J. S. Cardoso, “INbreast,” Academic 2018.
Radiology, vol. 19, no. 2, pp. 236–248, 2012. [88] R. K. Samala, H.-P. Chan, L. M. Hadjiiski, M. A. Helvie,
[72] J. Suckling, J. Parker, D. R. Dance et al., “The mammographic C. Richter, and K. Cha, “Evolutionary pruning of transfer
image analysis society digital mammogram database,” in learned deep convolutional neural network for breast cancer
Proceedings of the International Workshop on Digital diagnosis in digital breast tomosynthesis,” Physics in Med-
Mammography, Excerpta Medica International Congress icine & Biology, vol. 63, no. 9, article 095005, 2018.
Series, pp. 375–378, York, England, July 1994. [89] M. Benndorf, C. Herda, M. Langer, and E. Kotter, “Pro-
[73] N. Liu, J. Shen, M. Xu, D. Gan, E.-S. Qi, and B. Gao, vision of the DDSM mammography metadata in an ac-
“Improved cost-sensitive support vector machine classifier cessible format,” Medical Physics, vol. 41, no. 5, article
for breast cancer diagnosis,” Mathematical Problems in 051902, 2014.
Engineering, vol. 2018, Article ID 3875082, 13 pages, 2018. [90] M. D. Zeiler and R. Fergus, “Visualizing and understanding
[74] X. Liang, Z. Zhang, T. Niu et al., “Iterative image-domain convolutional networks,” in Proceedings of the European
ring artifact removal in cone-beam CT,” Physics in Medicine Conference on Computer Vision, vol. 1, pp. 818–833, Zurich,
& Biology, vol. 62, no. 13, pp. 5276–5292, 2017. Switzerland, September 2014.
[75] R. Zhang, W. Zhou, Y. Li, S. Yu, and Y. Xie, “Nonrigid [91] L. V. D. Maaten and G. Hinton, “Visualizing data using
registration of lung CT images based on tissue features,” t-SNE,” Journal of Machine Learning Research, vol. 9,
Computational and Mathematical Methods in Medicine, pp. 2579–2605, 2008.
vol. 2013, Article ID 834192, 7 pages, 2013. [92] N. Lei, Z. Luo, S.-T. Yau, and D. X. Gu, “Geometric un-
[76] H. Zhou, Y. Kuang, Z. Yu et al., “Image deformation with derstanding of deep learning,” 2018, https://arxiv.org/abs/
vector-field interpolation based on MRLS-TPS,” IEEE Access, 1805.10451.
vol. 6, pp. 75886–75898, 2018. [93] G. Montavon, W. Samek, and K.-R. Müller, “Methods for
[77] H. Greenspan, B. van Ginneken, and R. M. Summers, “Guest interpreting and understanding deep neural networks,”
editorial deep learning in medical imaging: overview and Digital Signal Processing, vol. 73, pp. 1–15, 2018.
future promise of an exciting new technique,” IEEE Trans- [94] L. Xing, E. A. Krupinski, and J. Cai, “Artificial intelligence
actions on Medical Imaging, vol. 35, no. 5, pp. 1153–1159, will soon change the landscape of medical physics research
2016. and practice,” Medical Physics, vol. 45, no. 5, pp. 1791–1793,
[78] H.-C. Shin, H. R. Roth, M. Gao et al., “Deep convolutional 2018.
neural networks for computer-aided detection: CNN ar- [95] Q. Huang, X. Huang, L. Liu, Y. Lin, X. Long, and X. Li, “A
chitectures, dataset characteristics and transfer learning,” case-oriented web-based training system for breast cancer
IEEE Transactions on Medical Imaging, vol. 35, no. 5, diagnosis,” Computer Methods and Programs in Biomedicine,
pp. 1285–1298, 2016. vol. 156, pp. 73–83, 2018.
[79] L. Zheng, Y. Zhao, S. Wang, J. Wang, and Q. Tian, “Good [96] D. Gu, C. Liang, and H. Zhao, “A case-based reasoning
practice in CNN feature transfer,” 2016, https://arxiv.org/ system based on weighted heterogeneous value distance
abs/1604.00133. metric for breast cancer diagnosis,” Artificial Intelligence in
[80] N. Tajbakhsh, J. Y. Shin, S. R. Gurudu et al., “Convolutional Medicine, vol. 77, pp. 31–47, 2017.
neural networks for medical image analysis: full training or [97] M. Jiang, S. Zhang, H. Li, and D. N. Metaxas, “Computer-
fine tuning?,” IEEE Transactions on Medical Imaging, vol. 35, aided diagnosis of mammographic masses using scalable
no. 5, pp. 1299–1312, 2016. image retrieval,” IEEE Transactions on Biomedical Engi-
[81] B. J. Erickson, P. Korfiatis, T. L. Kline, Z. Akkus, K. Philbrick, neering, vol. 62, no. 2, pp. 783–792, 2015.
and A. D. Weston, “Deep learning in radiology: does one size [98] K. Holland, A. Gubern-Mérida, R. M. Mann, and
fit all?,” Journal of the American College of Radiology, vol. 15, N. Karssemeijer, “Optimization of volumetric breast density
no. 3, pp. 521–526, 2018. estimation in digital mammograms,” Physics in Medicine and
[82] S. B. Kotsiantis, I. D. Zaharakis, and P. E. Pintelas, “Machine Biology, vol. 62, no. 9, pp. 3779–3797, 2017.
learning: a review of classification and combining tech- [99] J. Lee and R. M. Nishikawa, “Automated mammographic
niques,” Artificial Intelligence Review, vol. 26, no. 3, breast density estimation using a fully convolutional net-
pp. 159–190, 2006. work,” Medical Physics, vol. 45, no. 3, pp. 1178–1190, 2018.
[83] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How [100] R. Miotto, F. Wang, S. Wang, X. Jiang, and J. T. Dudley,
transferable are features in deep neural networks,” in Pro- “Deep learning for healthcare: review, opportunities and
ceedings of the Advances in Neural Information Processing challenges,” Briefings in Bioinformatics, vol. 19, no. 6,
Systems, pp. 3320–3328, Montréal, Canada, December 2014. pp. 1236–1246, 2017.
[101] Q. Dou, L. Yu, H. Chen et al., “3D deeply supervised network

for automated segmentation of volumetric medical images,”
Medical Image Analysis, vol. 41, pp. 40–54, 2017.
[102] J. Yun, J. Park, D. Yu et al., “Improvement of fully automated
airway segmentation on volumetric computed tomographic
images using a 2.5 dimensional convolutional neural net,”
Medical Image Analysis, vol. 51, pp. 13–20, 2019.
[103] G. Carneiro, J. Nascimento, and A. P. Bradley, “Unregistered
multiview mammogram analysis with pre-trained deep
learning models,” in Proceedings of the International Con-
ference on Medical Image Computing and Computer-Assisted
Intervention (MICCAI), pp. 652–660, Munich, Germany,
October 2015.
[104] N. Dhungel, G. Carneiro, and A. P. Bradley, “Fully auto-
mated classification of mammograms using deep residual
neural networks,” in Proceedings of the 2017 IEEE 14th In-
ternational Symposium on Biomedical Imaging (ISBI 2017),
pp. 310–314, Melbourne, Australia, April 2017.
[105] M. A. Al-masni, M. A. Al-antari, J. M. Park et al., “Detection
and classification of the breast abnormalities in digital
mammograms via regional convolutional neural network,”
in Proceedings of the Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC),
pp. 1230–1233, Jeju Island, South Korea, July 2017.
Hindawi
https://doi.org/10.1155/2019/1684218
Research Article
Automatic Segmentation of Pathological Glomerular Basement
Membrane in Transmission Electron Microscopy Images with
Random Forest Stacks
Lei Cao ,1 YanMeng Lu ,2 ChuangQuan Li,1 and Wei Yang 1
1
School of Biomedical Engineering, Southern Medical University, GuangZhou 510515, China
2
Central Laboratory, Southern Medical University, GuangZhou 510515, China
Correspondence should be addressed to Lei Cao; caolei@smu.edu.cn
Received 8 November 2018; Revised 2 February 2019; Accepted 24 February 2019; Published 25 March 2019
Copyright © 2019 Lei Cao et al. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Pathological classification through transmission electron microscopy (TEM) is essential for the diagnosis of certain nephropathy,
and the changes of thickness in glomerular basement membrane (GBM) and presence of immune complex deposits in GBM are
often used as diagnostic criteria. The automatic segmentation of the GBM on TEM images by computerized technology can
provide clinicians with clear information about glomerular ultrastructural lesions. The GBM region on the TEM image is not only
complicated and changeable in shape but also has a low contrast and wide distribution of grayscale. Consequently, extracting
image features and obtaining excellent segmentation results are difficult. To address this problem, we introduce a random forest-
(RF-) based machine learning method, namely, RF stacks (RFS), to realize automatic segmentation. Specifically, this work
proposes a two-level integrated RFS that is more complicated than a one-level integrated RF to improve accuracy and gen-
eralization performance. The integrated strategies include training integration and testing integration. Training integration can
derive a full-view RFS1 by simultaneously sampling several images of different grayscale ranges in the train phase. Testing
integration can derive a zoom-view RFS2 by separately sampling the images of different grayscale ranges and integrating the
results in the test phase. Experimental results illustrate that the proposed RFS can be used to automatically segment different
morphologies and gray-level basement membranes. Future study on GBM thickness measurement and deposit identification will
be based on this work.
1. Introduction of renal biopsies [4]. Therefore, TEM is essential for the

diagnosis of certain nephropathy. Considering the com-
Primary glomerular disease is the most common renal plexity of the TEM image of the glomerulus and related
disease in China [1]. The diagnosis of renal diseases is largely lesions, it is time consuming and labor intensive for a pa-
dependent on renal biopsy, which is regarded as the gold thologist to visually recognize subtle pathological changes,
standard. Transmission electron microscopy (TEM) com- resulting in a huge workload. Nevertheless, after the initial
bined with optical microscopy and immunofluorescence screening of the computer, the diagnostic efficiency and
examination constitutes a continuum of pathological di- accuracy of glomerular diseases can be improved with the
agnosis of renal diseases [2]. TEM allows the observation of help of automatic image-processing technology.
pathological changes in the microstructure of various glo- The diagnosis of many renal diseases is closely related to
merular cells that cannot be resolved under an optical the glomerular basement membrane (GBM) [5]. The
microscope. Thus observations from light and immune basement membrane, along with the lining of the endo-
pathology can be verified at an ultrastructure level [3]. theliocytes and the lining of the podocytes on the outside,
Studies have found that ultrastructural study provided forms the filtration barrier, allowing the blood to filter out
fundamental or important diagnostic information for 44.3% and form the urine, as shown in Figure 1. The changes of
Endotheliocyte
Podocyte
GBM
GV: 11-165 GV: 93-253

Figure 1: Interimage shape variations and intraimage grayscale inconsistency of GBM (GV: gray value).
thickness in GBM and presence of immune complex de- grayscale inconsistency. Figure 1 shows that the low contrast
posits in GBM are often used as diagnostic criteria for between the GBM segment and surrounding tissues, such as
certain nephropathy, such as membranous nephropathy endotheliocytes and podocytes, and variations in the form
with extensive membranous thickening and varying and width of GBM segments cause difficulty in autoex-
amounts of immune complexes, Alport syndrome with tracting features. In addition to the complex structure of a
diffuse membranous thickening, and familial recurrent pathological section, the grayscale distribution of TEM
hematuria syndrome (thin basement membrane ne- images is very wide because of the uncertainty of sample
phropathy) with diffuse thinning of GBM [6, 7]. Manual prefabrication and the illumination inhomogeneity of
measurement of the thickness of GBM is an early auxiliary transmission imaging.
[8, 9], but the workload is very expensive. Then, some To address the first challenge of autoextracting fea-
semiautomatic software tools [10, 11] are used to obtain tures, we employ a pixel-wise classifier, namely, random
the thickness of GBM more quickly and conveniently but forest (RF) [18], based on machine learning to avoid re-
still need manual intervention. In terms of morphological lying on hand-crafted features. RF is a committee of weak
complexity, the autoidentification difficulty of deposits is learners (e.g., decision tree) to solve classification and
the same as or even greater than GBM and literatures regression problems without manually specifying some
on this have not been found. In practicality, the thickness features through the construction and combination of
measurement and deposit identification can be realized multiple decision trees and random selection of attributes
subsequently and automatically on condition that [19, 20], which can be used to cope with the complex
the GBM region is completely autorecognized or is structural characteristics of biological images. RF has been
segmented. widely explored from medical image-processing fields,
Early in 1993, Ong et al. [12] applied adaptive window- especially detection tasks, including early identification or
based tracking to segment glomerular TEM images. Since prediction of Alzheimer’s disease [21], adrenal gland ab-
then, a few semiautomatic or fully automatic methods have normality detection [22], and automatic cardiac seg-
been proposed. Kamenetsky et al. [13] and Rangayyan et al. mentation [23].
[14] achieved GBM segmentation and measurement through An enhanced generalization effect based on a single RF
region division and dynamic contour modeling. Wu et al. classifier is hardly obtained because of the grayscale in-
[15] and Wu and Dikman [16] also proposed two methods. consistency intraimage. To address this second challenge,
One is to obtain the center line of the GBM by interpolating we propose an RF stack (RFS) model based on a wider
manual mark points and then autosegment GBM through grayscale range of images. After assigning TEM images to
distance mapping and low-pass filtering [15]. Another different grayscale groups, we sample from all these
method involves the use of threshold and morphological groups and train a full-view RF classifier as RFS1 and
method with no manual mark [16]. Liu et al. not only multiple RF zoom-view classifiers as RFS2 . In the seg-
segmented the GBM but also measured its length and mentation phase, each pixel of the new GBM image is
counted the number of slits [17]. Most existing methods classified automatically through full-view and zoom-view
mentioned above have made some contribution, but there RFS and the candidate segment results are combined and
are still many problems unresolved. These methods either optimized. Thus, the segmentation accuracy is improved
require tedious manual initialization that involves extra by using this two-level integrated machine learning
work for pathologists and introduces possible subjective method.
errors, or they can only be used to segment truncated GBM The remaining sections of this paper are organized
fragments with increased contrast and single direction as as follows. In Section 2, information regarding GBM
showed in their experimental results. Therefore, ensuring image selection and preprocessing is described and the
segmentation quality for the whole complex GBM images details of the proposed RFS model for GBM segmentation
remains challenging. is introduced. In Section 3, experimental results are re-
Two common difficulties associated with GBM seg- ported and discussed. In Section 4, the conclusion is
mentation are interimage shape variations and intraimage presented.
2. Materials and Methods

2.1. Data and Materials
2.1.1. Image Data. Renal biopsy specimens were immedi-

ately fixed with 2.5% cold glutaraldehyde with 0.1 M
phosphate buffer at pH 7.3 for 4 h, washed with phosphate
buffer, postfixed with 1% osmium tetroxide in the same
buffer, dehydrated with a graded series of ethanol, and
embedded in Spurr resin. Ultrathin sections (70 nm) were
contrast enhanced with uranyl acetate and lead citrate and
examined using a Hitachi H-7500 electron microscope
(Tokyo, Japan) at 60 kV. All of the sections were imaged with
MORADA G3 (EMSIS Corporation of Japan) at 5000x
magnification. In the field of vision, a whole glomerulus,
including the glomerular capillaries and the basement
membrane, was selected. Continuous filming was conducted (a) (b)
using the attached digital imaging system and controlled by Figure 2: TEM image (a) and the corresponding binary mask
the Pathological Image Workstation of the NanFang Hos- image of GBM (b).
pital in Guangzhou, China.
an iterative refinement scheme. The final segmentation is
2.1.2. Preprocessing of TEM Images. The pathologists col- selected from R1 and R2 by a human expert.
lected 351 images from the obtained glomerular TEM images
to build a GBM image database. Of these 351 images, 330
2.2.2. Software Tools. The image-processing and analysis
were used as a training set and divided into different groups
software FIJI (ImageJ) is developed by the US National
(N � 37) according to the range of GBM’s intensity. In the
Health Administration, and FIJI-based secondary develop-
test set, 21 images with different sizes and various basement
ment is well known. In this paper, we selected a FIJI plug-in
membrane types, such as stripe-, closed-, and compound-
named Trainable Weka Segmentation (TWS) (http://imagej.
type, were used. The image pixels were divided into GBM
net/Trainable_Weka_Segmentation) [24], which is based on
and background. Figure 2(a) shows the original TEM image,
the free open-source software Weka [25]. TWS combines a
and Figure 2(b) illustrates its corresponding GBM binary
series of machine learning algorithms to perform pixel-
mask manually labeled by the pathologist.
based image segmentation. Figure 4 shows that TWS is
modified and integrated with some image-processing
functions provided by Matlab to meet the needs of GBM
2.2. Method Overview
image segmentation.
2.2.1. Workflow Diagram. RF is an ensemble machine
learning method, which can be applied to image segmen-
2.3. Training
tation by classifying pixels into target or background. The
proposed RFS is a RF-based multilevel integrated structure 2.3.1. Random Forest. RF [23] is a common method for
that mainly involves two phases: hierarchical training and ensemble learning whose training algorithm relies on
refinement testing. The implementation process of RFS is bagging integration and random attribute selection in the
shown in Figure 3. construction of the decision tree. The training of one RF is
Train phase: from the prebuilt GBM database, an image shown in a blue arrow line in Figure 3. Bootstrap sampling
is randomly selected from each image group with different technology is used to generate T training subsets from the
GBM grayscale ranges. Hence, N images, namely, Img1 to original training set, and T decision tree models are
ImgN , needed for one training session are ready. The si- established to form one RF. An RF segmentation is illus-
multaneous sampling of N images and follow-up training trated in an orange arrow line in Figure 3. The test image is
yield RFS1 called the full-view RFS. Then, N RF classifiers, separately classified by the T decision trees in the RF, and
RFi (i � 1, . . . , N), are generated by sampling and training the result of each decision tree is aggregated to the final
each image individually, and RFS2 , which is called zoom- output by voting. In this paper, we refer this to level 1
view RFS, covering different grayscale ranges is integration.
constructed.
Test phase: each pixel of the test image is classified by
RFS1 and constitutes a candidate segmentation R1 . Each 2.3.2. RFS Classifiers. Considering that the intensities of
pixel of the test image is classified by each RFi in RFS2 and GBM in various TEM images are significantly different, a
got N coarse segmentation results, namely, CR1 to CRN . single RF classifier cannot extract different grayscale features
Then, another candidate segmentation R2 is obtained after of all TEM images and the segmentation performance is
Testing
Full-view
N ∗M R1
RFS1
Refinement
Img1 M RF1 CR1 Expert Final result

decision
Img2 M RF2 CR2
Voting
R2
Imgn M RFn CRn
Training
ImgN M RFN CRN
Zoom-view
RFS2
M M sampling and RF training
N ∗M
N ∗ M sampling AND RF training
RF segmentation
Figure 3: Flowchart of the proposed RFS method.
RFS system
Random forest Postprocessing
algorithms algorithms
TWS
Image-processing Machine learning

functions algorithms
FIJI Weka Matlab
Figure 4: Software tools.
unstable. For example, given an RF classifier sampled and methods were proposed. The full-view method simulta-
trained from Figure 2, the membrane illustrated in neously takes samples from N different grayscale range
Figure 5(a) can be well segmented as shown in Figure 5(b) images. As M pixels are sampled in each image, M × N
because of the similar grayscale of the GBM in Figure 5(a) pixels are sampled at the same time for training and RFS1 is
and the classifier. A poor result is obtained by using the same obtained. RFS1 is a large file in the same logic form of an RF.
classifier to segment the membrane in Figure 5(c), and the Due to the large number of sampled pixels and the limited
entire membrane fragment is almost not segmented as depth of the tree, we can assume that leaf nodes form
shown in Figure 5(d). stacks.
To address this problem, we introduced level 2 in- The zoom-view method separately takes samples from N
tegration. We first assigned all training images to N groups grayscale range images for training and obtains a series of
according to the average intensity in the GBM regions. RFs ranging from RF1 to RFN , thereby forming RFS2 . Since
Then, RFSs were constructed. Full-view and zoom-view the sampling points of each forest in RFS2 are well targeted
(a) (b)
(c) (d)
Figure 5: Segmentation results of TEM images with different grayscale ranges by using the same RF classifier.
to images with the similar range, the generalization per- map with the N coarse segmentation results, namely, CR1 to
formance of a single forest is limited and the results need to CRN :
be integrated and refined at the test or segment phase.
n(i.j)
p(i, j) � , (1)
N
2.3.3. Implementation Details. In the prebuilt GBM data-
base, the TEM images in the training set were divided into where N � 37 is the total number of coarse segmentation
N � 37 groups according to the average intensity of GBM. results, n(i, j) is the frequency of the pixel of ith row, jth
The average intensities of the GBM begin from about 73 column is marked as GBM by each CR, and p(i, j) is the
Hounsfield units with an intensity step of 3 Hounsfield units. probability of the pixel (i, j) as GBM.
TWS has 15 applicable features, and 14 of them were
selected as the inputs of the decision tree in this paper:
2.4.2. Postprocessing. In the probability map, a large gray
common grayscale features (mean, minimum, maximum,
value of a pixel corresponds to a high probability to become
median, and variance), boundary features (Sobel filter, Hes-
GBM. Therefore, by maximizing the similarity belonging to
sian, and difference of Gaussians), texture features (Gaussian
the same category or avoiding it to reach the minimum, the
blur, entropy, and Kuwahara filter), and other features
fuzzy C-means (FCM) [26] algorithm is utilized for post-
(membrane projections, Lipschitz filter, and neighbors).
processing to divide the image into the GBM regions and the
Other RF training parameters include the number of
background. Then, after some false positives are removed
decision trees (T � 100), the depth of decision trees
through a morphological operation, the GBM regions can be
(D � 9), and the number of sampling points per image
extracted.
(M � 2000). The selection of some parameters is discussed
in Section 4.
2.4.3. Iterative Refinement. Not every coarse segmentation
results CRi (i � 1, . . . , N) obtained from RFS2 provides
2.4. Segmentation. Given an image to be segmented or useful information on the construction of the probability
tested, two candidate segmentation results can be separately graph. In some extreme cases, some coarse segmentation
obtained by RFS1 and RFS2 . Pixel-by-pixel classification results are counterproductive to the probability map.
through full-view RFS1 yields a candidate segmentation R1 . Therefore, a refinement process is described in Algorithm 1.
The process of getting candidate R2 is more complicated. Figure 6 shows the whole segmentation process of RFS2 ,
After the preparation of a series of coarse results, namely, including (1) 37 RF classifiers, (2) probability map, (3)
CR1 to CRN , dealt by zoom-view RFS2 , the probability map postprocessing, and (4) iterative refinement.
are reconstructed and the candidate R2 can be obtained after
postprocessing and iterative refinement is completed. Fi-
nally, R1 and R2 are evaluated by experts to determine an 2.4.4. Manual Interaction for Final Decision. For a test
enhanced segmentation. image, the candidate segmentation results R1 and R2 are
compared with a gold standard labeled by a pathologist. For
new images without gold standard, a user can compare R1
2.4.1. Probability Map. For each image pixel to be seg- with R2 and make the final decision based on the following
mented, equation (1) is used to reconstruct the probability aspects:
(1) Coarse segmentation results with an obvious error GBM area are eliminated on the basis of the preset threshold w1 .
(2) The remaining K1 coarse segmentation results CRi (i � 1, . . . , K1 ) are used to obtain the probability map P1 and the binary mask
B1 as shown in Sections 2.4.1 and 2.4.2.
(3) The Jaccard similarity Si (i � 1, . . . , K1 ) of B1 and each CRi (i � 1, . . . , K1 ) are calculated. If the similarity Si is less than a preset
threshold w2 , the corresponding CRi is abandoned. The remaining K2 segmentation results are used to obtain a new probability
map P2 and the binary mask B2 .
(4) The Jaccard similarity (J) of B1 and B2 is calculated. If J ≤ 98%, let P1 � P2 and B1 � B2 , and step (3) is repeated until J > 98%
before they exit the loop.
(5) After the loop ends, the candidate result R2 equal to B2 is obtained.
ALGORITHM 1
CR1 CR2 CR3

. . . Iterative
Classifier p (i, j) = n (i, j)/N Postprocessing refinement
1-37
CR15 CR16 CR17

. . .
CR35 CR36 CR37
Figure 6: Segmentation process of RFS2 .
(1) Whether the foot process of the epithelial cell or the 3.1. Validation. The RFS method provides robust segmen-
cytoplasm of the endothelial cell is inappropriately tation results of GBMs with different morphologies and
contained in the region of the basement membrane grayscale ranges. RFS1 and RFS2 are trained with M � 2000
because its electron density is similar to that of the and N � 37. Figure 7 shows the segmented images obtained
basement membrane from the strip-, closed-, and compound-type basement
(2) Whether the subepithelial immune deposit is erro- membranes by using the RFS method. The top line shows
neously excluded from the basement membrane original images, and the bottom line shows the corresponding
because its electron density is higher than that of segmentation results. As can be seen from the figure, although
basement membrane the orientation, width, and other morphologies of the GBM
vary greatly, the results of the segmentation are relatively
(3) The continuity of the basement membrane should be
accurate. For the three test images shown, the Jaccard values
cautiously analyzed because pathological fracture
are higher than 0.75.
defects of the basement membrane are few
Figure 8 shows the segmented images from different
grayscale range basement membranes. It can be seen that
2.5. Evaluation Metrics. The accuracy of the proposed most of the GBM is accurately segmented. Compared with
method is evaluated by Jaccard coefficient, which is widely RF, the segmentation results of RFS are better. For example,
utilized to evaluate the performance of segmentation the original TEM image shown in Figure 5(c) fails with RF
methods [27–29]. It is a measure of geometric similarity segmentation, but it can be well segmented by RFS, as shown
defined by in column 1 of Figure 8.
As shown in Figures 7 and 8, although the morphology
A∩B and grayscale range of the basement membrane vary greatly,
Jaccard(A, B) � , (2)
A∪B the results of the RFS segmentation are stable, indicating a
good generalization performance of the RFS method. Future
where A and B are the results of manual segmentation by study on GBM thickness measurement and deposit iden-
human experts and the proposed method. The range of tification will be based on this.
Jaccard value is [0, 1]. A large coefficient value corresponds
to an accurate segmentation result.
3.2. Influence of the RF Classifiers of Different Grayscale
3. Experiments and Results Ranges. A multilevel integrated RFS classifier is constructed
to address the generalization problem of GBM segmenta-
In this study, 21 TEM images with different grayscale ranges, tion. This is based on the hypothesis that, for an RF classifier,
sizes, and basement membrane morphologies are used for the closer the grayscale range of the image to be segmented is
evaluation. These images are manually segmented by pa- to the training image, the better the segmentation effect will
thologists as the gold standard. be. The experimental results from the heat map in Figure 9
(a) (b) (c)
Figure 7: Segmentation results of RFS with different morphologies of GBM. (a) Strip type: size, 217∗307; Jaccard, 0.75. (b) Close type: size,
150∗206; Jaccard, 0.84. (c) Compound type: size, 282∗367; Jaccard, 0.76.
confirm this hypothesis. The horizontal axis from low to RFS2 are given in Algorithm1, involving the setting of
high represents 37 RF classifiers corresponding to 37 thresholds w1 and w2 . The Jaccard values of the segmen-
grayscale ranges of the training images. The range of the 1st tation result for each test image (the pink dots) by using
grayscale is [73.5, 75.5], and the range of the 37th grayscale is RFS2 with three different parameter combinations are shown
[145.5, 147.5]. The vertical axis represents the grayscale in columns 1, 2, and 3 of Figure 10, where w1 � 1 indicates
mean value of the GBM of the 21 test images. The grayscale that all N � 37 coarse segmentation results are included in
mean value of test image 1 is 78, which is close to the the steps of iterative optimization without filtering. The
grayscale range of levels 1 and 2. The 21st test image has an Jaccard values of the segmentation result by using full-view
average grayscale of 145, which is higher than the grayscale RFS1 are shown in column 4 of Figure 10.
range of level 37. Figure 10 shows that the full-view RFS1 is more robust
In this experiment, each test image is separately seg- than the zoom-view RFS2 . Regardless of how thresholds w1
mented by these 37 RF classifiers of different grayscale and w2 are set, the mean value of RFS2 is lower than RFS1 .
ranges, and the corresponding Jaccard value is shown in However, when w2 � 0.3, the segmentation result of some
different colors. An accurate segmentation result corre- test images of RFS2 is better than that of RFS1 . In the ex-
sponds to a high Jaccard value, and its color turns to bright perimental data, the maximum segmentation accuracy is
yellow. It can be seen from the color distribution in Figure 9 0.85 but the minimum value is almost 0 and all these values
that most RF classifiers are only sensitive to test images with are obtained through RFS2 .
a close grayscale range. However, as shown in Figures 7 and The stability of RFS1 is mainly caused by a large sample
8, the RFS can accurately segment test images with different training of the classifier, involving 2000 × 37 sample points.
grayscale ranges that means the generalization performance However, the disadvantage of this method is its high-
of RF is not good as proposed RFS. intensity computation. Some low accuracy of RFS2 is
caused by the effect of N course segmentation results on
refinement process. If the similarity of most results with the
3.3. Differences between Full-View RFS1 and Zoom-View RFS2 . gold standard is insufficient, a poor optimized image is
The multilevel RFS is constructed in full view and zoom view obtained. Otherwise, the result is enhanced or even exceed
as shown in Figure 3. The refinement details of zoom-view that of RFS1 . Therefore, the final step of this method involves
(a) (b) (c)
Figure 8: Segmentation results of RFS with various grayscale ranges of GBM (RoGV: range of gray value). (a) RoGV: 54–255; size: 282∗274;
Jaccard: 0.70. (b) RoGV: 43–187; size: 168∗308; Jaccard: 0.71. (c) RoGV: 12–167; size: 297∗408; Jaccard: 0.65.
Jaccard
157
139 0.8
134
134 0.7
134
131
129 0.6
Grayvalue of images
120
117 0.5
117
115
106 0.4
106
106 0.3
95
85
84 0.2
83
82 0.1
80
78 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
Grayscales
Figure 9: Heat map of segmentation results of different grayscale ranges of RF classifiers.
the selection between the two candidate segmentation re- variance of methods (1)–(4), full-view RFS1 (5), and the final
sults of RFS1 and RFS2 . result (6). The mean accuracy of V is the lowest among them.
V increases after FCM postprocessing and further improves
3.4. Effects of Postprocessing and Refinement. The following after iterative refinement is completed. The final result in-
methods are adopted to validate the effect of postprocessing cludes the maximum mean and a relatively small variance.
and iterative refinement on RFS2 : (1) voting method (V), (2)
additional postprocessing with FCM on the voting result 4. Discussion
(V + F), (3) additional iterative refinement on the voting result
(V + I), and (4) additional iterative refinement on method 2 4.1. Methods for Constructing Ensembles. Ensemble methods
(V + F + I), namely, RFS2 . Figure 11 shows the mean and construct a set of classifiers and then classify new data
0.9 usage. In this study, the number of decision trees is ex-

0.8
perimentally set to 100.
The number of sampling points is another critical pa-
0.7 rameter in the RFS method. In our experiment, as the
number of sampling points increases from 200 to 2000 per
0.6
training image, the accuracy rate of the RFS classifier in-
0.5 creases by approximately 10%, whereas the accuracy is not
Jaccard
improved greatly if the number of sampling points con-

0.4
tinuously increases. Thus, the total sampling point is set to
0.3 74,000, where M � 2000 and N � 37, to obtain the best
result.
0.2
In TWS, 15 available image feature attributes are pro-
0.1 vided in the decision tree construction. In our experiment,
the application of most features can improve the accuracy of
0
w1 = 1, w1 = 1, w1 = 0.3, Full view segmentation, but entropy (E) and anisotropic diffusion (A)
w2 = 0.03 w2 = 0.3 w2 = 0.3 are time consuming. Let 13F indicates 13 other features aside
Zoom view Zoom view Zoom view
from E and A; Figure 12 shows that the application of feature
Figure 10: Jaccard values of zoom-view RFS2 at three different A not only costs more time but also reduces the accuracy of
parameter combinations and full-view RFS1 . segmentation whether by using classifier RFS1 or RFS2 .
Therefore, only 14 features other than anisotropic diffusion
are used to construct the decision tree in the proposed RFS
0.8 method.
0.6
0.4 4.3. Limitations of the Proposed Method. Our experiment

results reveal that the proposed RFS method obtains poor
0.2 performance for some cases. For example, for a low-
contrast image, the accuracy rate of voting is almost 0.
0
Only the accuracy rate of V + I reaches 23%, whereas the
V V+F V+I V + F + I Full view Final
accuracy rates of the other methods are below 20%, even
Jaccard that of RFS1 . Such bad results greatly reduce the average
Figure 11: Effects of postprocessing and refinement methods. accuracy of the RFS method. RF can be combined with
other pattern recognition methods for better performance.
Lu et al. [32] applied incomplete RF with a robust vector
points by taking a weighted vote of their predictions. machine for the early identification of mild cognitive
Dietterich [30] assumed that five general purpose ensemble impairment. This method outperforms two other semi-
methods exist: enumeration of hypotheses, manipulation of supervised learning methods. Therefore, to improve the
training examples, manipulation of input features, ma- segmentation accuracy of low-contrast GBM images, the
nipulation of output targets, and injection of randomness. combination of RF methods with other methods will be our
We adopted two of them and developed their corre- future work.
sponding RFS methods. Full-view RFS2 manipulates the
training examples to generate multiple hypotheses. Con-
sidering the complicated GBM images, we sample multiple 5. Conclusion
grayscale images to further increase the diversity of the
hypotheses. Zoom-view RFS1 manipulates the output The segmentation of the whole GBM region in TEM
probability graph to achieve integration. The iterative re- pathological images can provide more rapid and intui-
finement step is addressed to reduce the adverse effect of tionistic observation for the morphological change and can
the rough segmentation result, further improving the ac- reduce the tedious and expensive manual workload of the
curacy of segmentation. pathologist. This work proposed a two-level integrated RFS
method involving training integration and testing in-
tegration to autosegment a GBM image. A total of 351
4.2. Selection of the Parameters. The number of decision clinical images were included in the experiment. The ac-
trees is among the most important parameter in the ap- curacy and generalization ability of the RFS method were
plication of RF algorithm in medical image segmentation validated. Experimental results illustrated that the pro-
[31]. Theoretically, with the increasing number of decision posed method could be used for the automatic segmen-
trees, the classification accuracy of the algorithm gradually tation of GBM with different morphological characteristics
increases as computational cost rapidly increases. The and grayscale ranges. Further study is underway to im-
optimal number of trees should obtain a good balance prove segmentation accuracy of the automated CAD
between evaluation metric, processing time, and memory system and to implement GBM thickness measurement
0.65 [6] G. Osawa, P. Kimmelstiel, and V. Sailing, “Thickness of

0.6 glomerular basement membranes,” American Journal of
Clinical Pathology, vol. 45, no. 1, pp. 7–20, 1966.
0.55
[7] S. Saxela, D. J. Davies, and R. L. G. Kirsner, “Thin basement
0.5
membranes in minimally abnormal glomeruli,” American
Jaccard
0.45 Journal of Clinical Pathology, vol. 43, pp. 32–38, 1990.

0.4 [8] J. G. Basta, V. S. Venkataseshan, J. Gil, D. U. Kim,
0.35
S. H. Dikman, and J. Churg, “Morphometric analysis of
glomerular basement membranes (GBM) in thin basement
0.3
membrane disease (TBMD),” Clinical Nephrology, vol. 33,
0.25 pp. 110–114, 1990.
13F 13F + E 13F + A 13F + E + A
[9] B. Marquez, I. Zouvani, A. Karagrigoriou et al., “A simplified
RFS1 method for measuring the thickness of glomerular basement
RFS2 membranes,” Ultrastructural Pathology, vol. 27, no. 6,
Figure 12: Effects of different image features with or without pp. 409–416, 2003.
entropy and anisotropic diffusion. [10] S. E. Thomson, S. V. McLennan, P. D. Kirwan et al., “Renal
connective tissue growth factor correlates with glomerular
basement membrane thickness and prospective albuminuria
and deposit autorecognition for auxiliary pathological in a non-human primate model of diabetes: possible pre-
diagnosis. dictive marker for incipient diabetic nephropathy,” Journal of
Diabetes and Its Complications, vol. 22, no. 4, pp. 284–294,
2008.
Data Availability [11] K. Kyriacou, M. Nearchou, I. Zouvani et al., “The many faces
of thin basement membrane nephropathy; a population based
The image data used to support the findings of this study are
study,” in Topics in Renal Biopsy and Pathology, pp. 93–114,
available from the corresponding author upon request. IntechOpen, London, UK, 2012.
[12] S. H. Ong, S. T. Giam, R. Sinniah et al., “Adaptive window-
Conflicts of Interest based tracking for the detection of membrane structures in
kidney electron micrographs,” Machine Vision and Applica-
The authors declare that there are no conflicts of interest tions, vol. 6, no. 4, pp. 215–223, 1993.
regarding the publication of this article. [13] I. Kamenetsky, R. M. Rangayyan, and H. Benediktsson,
“Analysis of the glomerular basement membrane in images of
renal biopsies using the split-and-merge method: a pilot
Acknowledgments study,” Journal of Digital Imaging, vol. 23, no. 4, pp. 463–474,
2010.
This work was supported in part by the grants from the [14] R. M. Rangayyan, I. Kamenetsky, and H. Benediktsson,
National Natural Science Foundation of China (no. “Segmentation and analysis of the glomerular basement
81771916) and the Guangdong Provincial Key Laboratory of membrane in renal biopsy samples using active contours: a
Medical Image Processing (no. 2014B030301042). Many pilot study,” Journal of Digital Imaging, vol. 23, no. 3,
thanks are due to Xi Yu, Yuanyuan Liao, Can Xu, Zhenxing pp. 323–331, 2010.
Li, and Zitao Zhang, who participated in this study during [15] H. S. Wu, S. Dikman, and J. Gil, “A semi-automatic algorithm
their graduation project in Southern Medical University and for measurement of basement membrane thickness in kidneys
made some contributions. in electron microscopy images,” Computer Methods and
Programs in Biomedicine, vol. 97, no. 3, pp. 223–231, 2010.
[16] H. S. Wu and S. Dikman, “Segmentation and thickness
References measurement of glomerular basement membranes from
[1] F.-D. Zhou, M.-H. Zhao, W.-Z. Zou et al., “The changing electron microscopy images,” Journal of Electron Microscopy,
spectrum of primary glomerular diseases within 15 years: a vol. 59, no. 5, pp. 409–418, 2010.
survey of 3331 patients in a single Chinese centre,” Nephrology [17] J. Liu, “Implementation of a semi-automatic tool for analysis
Dialysis Transplantation, vol. 24, no. 3, pp. 870–876, 2009. of TEM images of kidney samples,” Master Thesis, Uppsala
[2] M. Haas, “A reevaluation of routine electron microscopy in University, Uppsala, Sweden, 2012.
the examination of native renal biopsies,” Journal of the [18] L. Breiman, “Random forests,” Machine Learning, vol. 45,
American Society of Nephrology, vol. 8, no. 1, pp. 70–76, 1997. no. 1, pp. 5–32, 2001.
[3] M. Santostefano, F. Zanchelliet, A. Zaccaria et al., “The ul- [19] T. K. Ho, “The random subspace method for constructing
trastructural basis of renal pathology in monoclonal gam- decision forests,” IEEE Transactions on Pattern Analysis and
mopathies,” Journal of nephrology, vol. 18, no. 6, p. 659, 2005. Machine Intelligence, vol. 20, no. 8, pp. 832–844, 1998.
[4] M. Wagrowska-Danilewicz and M. Danilewicz, “Current [20] T. K. Ho, “Random decision forests,” in Proceedings of 3rd
position of electron microscopy in the diagnosis of glomerular International Conference on Document Analysis and Recog-
diseases,” Polish Journal of Pathology, vol. 58, no. 2, pp. 87–92, nition, vol. 1, p. 278, Montreal, Canada, August 1995.
2007. [21] A. V. Lebedev, E. Westman, G. J. P. VanWesten et al.,
[5] A. Sementilli, L. A. Moura, and M. F. Franco, “The role of “Random Forest ensembles for detection and prediction of
electron microscopy for the diagnosis of glomerulopathies,” Alzheimer’s disease with a good between-cohort robustness,”
Sao Paulo Medical Journal, vol. 122, no. 3, pp. 104–109, 2004. NeuroImage: Clinical, vol. 6, no. C, pp. 115–125, 2014.
[22] G. Saiprasad, C. I. Chang, N. Safdar et al., “Adrenal gland

abnormality detection using random forest classification,”
Journal of digital imaging, vol. 26, no. 5, pp. 891–897, 2013.
[23] D. Mahapatra, “Automatic cardiac segmentation using se-
mantic information from random forests,” Journal of Digital
Imaging, vol. 27, no. 6, pp. 794–804, 2014.
[24] I. Arganda-Carreras, V. Kaynig, C. Rueden et al., “Trainable
Weka Segmentation: a machine learning tool for microscopy
pixel classification,” Bioinformatics, vol. 33, no. 15, pp. 75–80,
2017.
[25] E. Frank, M. Hall, G. Holmes et al., “Weka-a machine learning
workbench for data mining,” in Data Mining and Knowledge
Discovery Handbook, pp. 1269–1277, 2005.
[26] J. Nayak, B. Naik, and H. S. Behera, “Fuzzy C-means (FCM)
clustering algorithm: a decade review from 2000 to 2014,” in
Smart Innovation, Systems and Technologies, vol. 32,
pp. 133–149, Springer, New Delhi, India, 2015.
[27] A. Padma and R. Sukanesh, “Segmentation and classification
of brain CT images using combined wavelet statistical texture
features,” Arabian Journal for Science and Engineering, vol. 39,
no. 2, pp. 767–776, 2014.
[28] B. Tóth, F. Lillo, and J. D. Farmer, “Segmentation algorithm
for non-stationary compound Poisson processes,” The
European Physical Journal B, vol. 78, no. 2, pp. 235–243,
2010.
[29] N. M. Noor, J. C. M. Than, O. M. Rijal et al., “Automatic lung
segmentation using control feedback system: morphology and
texture paradigm,” Journal of Medical Systems, vol. 39, no. 3,
p. 214, 2015.
[30] T. G. Dietterich, “Ensemble Methods in Machine Learning,”
in Multiple Classifier Systems, LNCS 1857, pp. 1–15, Springer,
Berlin, Germany, 2000.
[31] T. M. Oshiro, P. S. Perez, and J. A. Baranauskas, “How many
trees in a random forest?,”in Machine Learning and Data
Mining in Pattern Recognition. Lecture Notes in Computer
Science, vol. 7376, no. 4, pp. 154–168, Springer, Berlin,
Germany, 2012.
[32] S. Lu, Y. Xia, W. Cai, M. Fulham, and D. D. Feng, “Early
identification of mild cognitive impairment using incomplete
random forest-robust support vector machine and FDG-PET
imaging,” Computerized Medical Imaging and Graphics,
vol. 60, pp. 35–41, 2017.
Hindawi
https://doi.org/10.1155/2019/9079840
Research Article
Minimalistic Approach to Coreference Resolution in
Lithuanian Medical Records
Voldemaras Žitkus,1 Rita Butkienė,1 Rimantas Butleris,1 Rytis Maskeli�

unas ,1
Robertas Damaševičius ,1 and Marcin Woźniak 2
1
Faculty of Informatics, Kaunas University of Technology, 51386 Kaunas, Lithuania
2
Institute of Mathematics, Silesian University of Technology, 44-100 Gliwice, Poland
Correspondence should be addressed to Robertas Damaševičius; robertas.damasevicius@ktu.lt
Received 18 January 2019; Accepted 26 February 2019; Published 20 March 2019
Guest Editor: Ka L. Man
Copyright © 2019 Voldemaras Žitkus et al. This is an open access article distributed under the Creative Commons Attribution License,
Coreference resolution is a challenging part of natural language processing (NLP) with applications in machine translation,
semantic search and other information retrieval, and decision support systems. Coreference resolution requires linguistic
preprocessing and rich language resources for automatically identifying and resolving such expressions. Many rarer and under-
resourced languages (such as Lithuanian) lack the required language resources and tools. We present a method for coreference
resolution in Lithuanian language and its application for processing e-health records from a hospital reception. Our novelty is the
ability to process coreferences with minimal linguistic resources, which is important in linguistic applications for rare and
endangered languages. The experimental results show that coreference resolution is applicable to the development of NLP-
powered online healthcare services in Lithuania.
1. Introduction coreference resolution and semantic annotation. The main

issue here is the ambiguity and complexity of the natural
Digital means of medical informatics, especially when ap- language, thus making the progress in IE dependent on the
plying natural language processing (NLP), are indispensable evolution of the NLP techniques. While for widely used
in the application of e-health and digitalization of medical languages (such as English), the IE-related NLP research has
records and processes [1]. The use of NLP has proved as a already reached the levels of maturity and practical appli-
lower-cost alternative to traditional medical methods in cation on a massive scale (e.g., IBM Watson project) [6], but
many cases such as to forecast stress symptoms and suicide the resource-poor languages, such as Lithuanian [7], remain
risk in free-text responses sent via a mobile phone [2], or to an open NLP research field. The baseline application is often
detect seasonal disease outbreaks by monitoring search steered towards automated concept extraction [8, 9], often in
engine queries [3], and discovery of healthcare knowledge combination with text mining [10, 11].
from social media [4, 5]. NER when applied to biomedical texts is a critical step for
With the development of Semantic Web technology, web developing decision support tools for smart healthcare. Ex-
information retrieval (IR) is changing towards meaning- amples for it are as follows: drug name recognition (DNR),
based IR. The quality of retrieved documents relevant to the which recognizes pharmacological substances from bio-
user also highly depends on the information extraction (IE) medical texts and classifies them for discovering drug-drug
methods applied. In general, IE focuses on automatic ex- interactions [12, 13]; biomedical named entity recognition
traction of structured information from the unstructured (BNER), which extracts biomedical concepts of interest such
source. Standard document text preprocessing steps used in as genes and proteins [14]; and medical entity recognition,
IE are lexical analysis, morphological analysis, and named which is information extraction from unstructured electronic
entity recognition (NER), which can be complemented by health records [15–17]. Such studies include mapping clinical
descriptions to Systematized Nomenclature of Medicine based on a set of statistical measures and syntactical and
codes [18] or other medical lexicons. Unstructured texts in the semantic information [28]. “Off the shelf” type of IR al-
medical domain contain valuable medical information, and gorithms can be utilized quite successfully in some of the
there are many errors, such as spelling errors, improper scenarios, especially with limited focus areas (in a medical
grammatical use, and semantic ambiguities, which hinder sense) [29]. Accuracy can be further improved by analysing
data processing and analysis [19]. Structuration of medical trigram frequencies [30] and applying graph-style algo-
domain knowledge using biomedical ontologies and con- rithms [31] in context-sensitive corpus fragments.
trolled vocabularies provide support for data standardization In general, the coreference resolution methods can be
and interoperability, healthcare administration, and clinical classified into knowledge rich and knowledge poor. Both
decision support [20]. Rich concepts linked by semantic methods require large resources such as semantic in-
relationships such as in the Unified Medical Language System formation, syntactic annotations, or preannotated corpora
(UMLS) contribute to healthcare data integration, pattern of hospital transcripts from a hospital reception. Under
mining from EHRs, medical entity recognition in clinical text, resourced, rarer languages, like Lithuanian, usually do not
and clinical data sharing [21]. The development of online have such resources available.
healthcare services as powerful platforms provides users with While up-to-date, no research has been performed to
an opportunity to address health concerns such as improving solve coreferences in Lithuanian, but many solutions have
patient-centered care and supporting self-management. The been proposed for other languages, mostly for English
users consider online healthcare services as a vital source of (Table 1). Note that the evaluation results are not directly
health information but still need more powerful semantic comparable, as the authors used different corpora.
search engines to arrive at informed decisions on their own Considering languages that are more-or-less grammat-
health and for more active participation in healthcare pro- ically similar to Lithuanian (which is one of the Baltic
cesses [22]. However, these services depend greatly upon the languages), we summarize the related work on Latvian (only
support provided by the natural language processing. other Baltic language) and Slavic languages such as Polish,
Our previous work included the development of the Russian, and Czech in Table 2.
semantic search framework [23] for answering questions
(i) For Latvian, the only solution is LVCoref [45]. It is a
presented in structured Lithuanian language, which is based
rule-based system that uses an entity-centric model.
on Semantics of Business Vocabulary and Business Rules
It focuses on named entity matches (exact matches,
(SBVR) language. Our results in [23] showed that there is a
acronyms) and uses Hobbs’ algorithm for pronouns.
strong need to complement the NLP pipeline of semantic
search with the coreference resolution. Such coreference (ii) For Polish, rule-based Ruler [46] for scoring of can-
resolution tools have not been developed for the Lithuanian didates uses coreferences gender/number and in-
language yet, especially for very sector-important digital cluding (removal of nested groups) rules, lemma, and
medical application, as we apply our algorithm to process Wordnet rules for nominal expressions and pronoun
digital transcripts of a hospital reception. Therefore, the rule specifically targeting pronouns. BARTEK [47] is
creation of such tools is a prerequisite for further im- an adaptation of BART, which was designed for
provement of NLP-supported health-oriented decision English, to Polish. Mixed Polish coreferences resolu-
making in the Lithuanian language, while the experience tion approach combines neural networks architecture
gained could be extended to developing semantic search with the sieve-based approach [48].
tools for other under-resourced languages as well. (iii) For Russian, RU-EVAL-2014 [49] was an evaluation
The rest of the paper is structured as follows. In Section 2, campaign of anaphora and coreferences resolution
we analyse the related works in the coreference resolution tools that employed a wide variety of approaches.
field, including the state-of-the-art and the methods pro- The evaluation was performed on Russian Cor-
posed for languages which are grammatically similar to eference Corpus (RuCur). Machine learning ap-
Lithuanian. Sections 3 and 4 present the coreference reso- proaches [50] were also used.
lution algorithm and its experimental evaluation. Section 5 (iv) For Czech, coreferences are annotated in the tec-
presents conclusions and discusses future work. togrammatical layer of Prague Dependency Tree-
bank (PDT) and their first coreference resolution
2. Related Works approach was rule based [51]. At first, all possible
candidates are collected and then their list is nar-
Machine-learning and rule-based approaches are efficient rowed down using 8 filters, and then from
methods in semantic processing, especially when enhanced remaining ones closest to corefering object is se-
with external knowledge and coreference clues derived from lected as antecedent. Nguy et al. [52] adapted two
the structured document, while often still performing better older English language approaches to Czech lan-
(in comparison with classic implementations) in coreference guage and used Decision Tree C5 for the classifier-
resolution when provided with ground truth mentions [24], based approach, while the ranker-based approach
while further expanded with scaffolding approaches [25]. employed the averaged perceptron algorithm.
Unsupervised methods can be applied to large-scale sce- Both approaches were trained and evaluated on
narios [26, 27]. Alternatively, a hybrid strategy may be used PTD data with ranker-based approach providing
Table 1: Comparison of coreference resolution approaches.

Method Foundation Precision Recall F1
Syntactic tree with labeled nodes, syntactic rules,
Hobbs [32] 0.81–0.91 na na
selection constraint rules
BFP [33] Centering theory 0.49–0.90 na na
Left-right centering [34] Modified centering theory 0.72–0.81 na na
Mitkov [35] POS tagger, antecedent indicators 0.897 na na
RAP [36] Salience factors 0.85–0.89 na na
Xrenner [37] Syntactic and semantic rules 0.51–0.55 0.49–0.57 0.49–0.56
Probabilistic [38] Bayesian rule 0.82–0.84 na na
MARS [39] Genetic algorithms 0.53–0.84 na na
Soon et al. [40] Machine learning (decision tree C5) 0.65–0.69 0.53–0.56 0.62
ILP [41] Machine learning (logistic classifier) 0.78–0.89 0.47–0.58 0.61–0.68
Wiseman et al. [42] Deep learning 0.77 0.70 0.73
Lee et al. [43] Deep learning 0.81 0.73 0.77
Žitnik et al. [44] Conditional random fields 0.68–0.94 0.30–0.87 0.41–0.87
Table 2: Comparison of coreference resolution methods for Balto-Slavic languages.

Method Foundation Precision Recall F1
LVCoref [45] Rule based, Hobbs’ algorithm 0.69–0.88 0.66–0.80 0.68–0.84
Ruler [46] Rule based 0.59–0.65 0.50–0.75 0.55–0.69
BARTEK [47] Machine learning 0.58 0.65 0.61
Mixed [48] Deep learning, sieve based 0.70 0.68 0.69
RU-sys1 [49] Rule based, ontology 0.82 0.70 0.76
RU-sys2 [49] Rule based 0.71 0.58 0.64
RU-sys3 [49] Rule based 0.63 0.50 0.55
RU-sys4 [49] Statistical, ontology 0.54 0.51 0.53
RU-sys5 [49] Machine learning, semantics 0.58 0.42 0.49
RU-sys6 [49] Decision tree 0.36 0.15 0.21
Khadzhiiskaia and Sysoev [50] Machine learning 0.84 0.77 0.80
Kučová and Žabokrtský [51] Rule-based filters 0.60 na na
CZ classifier [52] Classifier-based machine learning 0.70–0.76 0.70–0.76 0.70–0.76
CZ ranker [52] Ranker-based machine learning 0.79 0.79 0.79
Treex CR (Czech, English) [53] Machine learning na na 0.61–0.68
Treex CR (Russian, German) [54] Machine learning, projection 0.50–0.64 0.15–0.24 0.25–0.34
better results. Treex CR [53] was developed for the text (antecedent). For example, “Tom has a backache. He was
Czech language and adapted to English, Russian, injured.” Here the words “Tom” and “He” refer to the same
and German, although for Russian and German, entity. Without resolving the relationship between these two
English coreferences labels were projected, which structures, it would not be possible to determine why Tom
produced notably lower results [54]. has the backache, nor who was injured. In such cases, se-
mantic information would be lost.
In summary, the rule-based solutions have the advantages
Anaphoric objects are expressed with pronouns and
of easier adaptability and provide comparable results when
cannot be independently interpreted without going back to
good training data are not available as is the case for Lithuania.
its antecedent. In this work, such expressions are called
Many of more advanced solutions cannot be fully adapted for
coreferences, unless it is required to make a distinction.
rarer and under-resourced languages due to the lack of
Usage of such expressions can vary depending on the type
available linguistic resources, as is the case with Lithuanian
and the style of the text. Here we focus on texts from
language. For example, BARTat the time supported 64 feature
medical-related domains.
extractors, but due to lack of language-specific resources for
The role of coreference resolution in the semantic search
the Polish language, only 13 could be utilized. The solutions
framework is to provide additional semantic information
that are not heavy on linguistic resources can be very useful for
after named entity recognition before semantic annotation
resource-poor languages in general.
(Figure 1).
3. A Rule-Based Coreference Resolution: A
Lithuanian Case
3.2. Conceptual Model of Coreference Resolution. In this
3.1. Definition and Framework. Coreference resolution (or chapter, a conceptualization of coreference resolution is
anaphora) is an expression, the interpretation of which presented. A given model, which is expressed as UML class
depends on another word or phrase presented earlier in the diagram (Figure 2), specifies the concepts playing a certain
NLP pipeline
Morphological Named entity Coreference Semantic

Lexical analysis
analysis recognition resolution annotation
A-box
assertations
Document Stand-off annotations
Figure 1: An NLP pipeline of semantic search framework with coreference resolution component.
Concepts of Input Flow Concepts of Output Flow

Text
1 (i) publication_date 1 resolved_in
1
consists_of 0..1
1..∗ (i) previous
follows
Lexical_unit
mentioned_in (i) value

(ii) start_position 0..1
(iii) length (i) next
0..*
includes
Sentence Lexeme fits Coreference
1 1 contains 1..∗
(i) lemma 1..∗ 0..1 (i) type: Ref_Type
(ii) pos: POS_Type (ii) subtype: R_Subtype
Named_Entity
expressed_by (iii) gender: G_Type (iii) position: P_Type
0..∗ (i) start_position (iv) number: N_Type (iv) group: Gr_Type
(ii) length 0..1 1..∗ (v) start_position
0..∗ 1..∗ (vi) length
0..1 0..1 fits
Person_NE 1 0..1
0..∗ (i) gender: G_Type Other_Part_Of_ Pronoun Noun refers_to
1..∗
Speech (i) type: Pr_Type Mention
Organization_NE
0..∗ (i) start_position
0..∗ (ii) length
Location_NE Comma Conjunction express
0..∗
fits
mentioned_us fits
Database of Public Persons Classification of Professions

0..1 0..1
holds describes (i) hyponym
Known_Person Position_Held Profession_Name Profession
0..∗
(i) full name 1 0..∗ (i) position_name (i) value 1..∗ 1
(ii) gender: G_Type (ii) lemma (ii) lemma
-hypernym 1
(iii) from_date
(iv) to_date
broadnes
≪enumeration≫ ≪enumeration≫ ≪enumeration≫ ≪enumeration≫ ≪enumeration≫ ≪enumeration≫ ≪enumeration≫ ≪enumeration≫

Ref_Type R_Subtype P_Type Gr_Type POS_Type Pr_Type G_Type N_Type
Nominal General Irrelevant Single Noun Demonstrative Male Singular

Pronominal Relative Backward Multiple Pronoun ... Female Plural
... Repetition Forward Ambiguous Preposition
Hypernym_Hyponym Separator
Synonym ...
Feature
Figure 2: A conceptual model of a coreference resolution domain.

role in the extraction of coreferences of a certain type. The in case of repetitions), and group (is singular, refers to the
model gives us an understanding of the following: coreference group or is ambiguous) are specified. Each
referent refers at least to one coreferent (a concept Mention).
(i) What features of text, sentence, and word help us
Each Mention starts at a certain position in the text, is of a
recognize the existence of coreference (they are
certain length, and fits at least one Lexeme. Some of them can
specified in the package Concepts of Input Flow)
fit a certain Named_Entity.
(ii) What kind of text preprocessing is required
(iii) What additional resources are required for reso-
3.3. Coreference Resolution Algorithms. The decision table
lution of certain type coreferences (they are speci-
with guidelines for the application of the certain resolution
fied in the package Database of Public Persons and
algorithm is shown as Figure 3. The conditions are checked
Classification of Professions)
consecutively on every lexeme in the text, and, if the con-
For example, from the model provided, it is clear that, dition is satisfied, a corresponding algorithm is activated.
before coreference resolution starts, it is important to pre- For example, if C2 condition is met then immediately A1
process text and obtain the following: algorithm is activated.
For resolution of a specific type of references, we propose
(i) A text segmented into sentences and lexemes
the following algorithms:
(ii) Morphological features of lexemes identified
(i) A1: specific rules resolution algorithm for resolution
(iii) Named entities recognized
of certain usage of pronouns
Text preprocessing itself is not a task of coreference (ii) A2: general pronoun resolution algorithm which
resolution, so it is out of the scope of this paper. focuses on the cases where pronouns refer to nouns
It is worthy to mention that the model is quite abstract, (or noun phrases) that are recognized as named
language independent, and technology independent. There- entities of “person” class
fore, it is applicable not only for Lithuanian but for gram-
(iii) A3: PRA (partial, repetition, and acronym) reso-
matically similar languages as well. Concepts of this model are
lution algorithm for resolution of nouns recognized
used for the formalization of coreference resolution rules in the
as named entities and their repeated usage in the
next section. The concepts are explained in more detail below.
same text
The main concepts of coreference resolution are Text,
Lexical_Unit, and Named_Entity. The concept Text assumes a (iv) A4: HHS (hypernym, hyponym, synonymous)
textual document whose content should be analysed. Each test resolution algorithm for resolution of nouns rec-
has an associated publication date, which is important for ognized as profession names including their syno-
solving coreferences. Each text consists of at least one Lex- nyms and hypernyms/hyponyms
ical_Unit, which includes paragraphs, sentences, words, and (v) A5: feature resolution algorithm for resolution of
punctuations, classified into the Sentence and Lexeme cate- nouns that represent certain feature (at the moment
gories. Lexeme assumes lexical units such as words, punctu- only public position being held) of the named entity
ations, and numbers. Each lexeme is characterized by a lemma of a person
and a part of speech, and some of them (nouns and pronouns)
The coreference resolution starts from the sequential
by grammatical gender and number. The lexeme could be
analysis of each lexeme looking for a certain type of pronoun
specialized by POS category: Noun, Pronoun, and Oth-
and noun. Depending on identified features of lexeme, a
er_Part_Of_Speech. Special cases of Other_Part_Of_Speech are
decision about further analysis is taken. The decision table
Comma and Conjunction, which are required for the de-
(Figure 3) summarizes conditions for the application of the
scription of conditions of some coreference resolution rules.
certain resolution algorithm. The conditions are listed in the
A Named_Entity concept defines an object to whom
upper left quadrant; the decision alternatives are listed in the
pronouns or certain nouns can refer. NER algorithms
lower left quadrant. The upper right quadrant shows the
usually recognize three types of entities: a person (Per-
possible alternatives for the conditions of the corresponding
son_NE), an organization (Organization_NE), and a location
row. In the upper right quadrant, the answer “na” stands for
(Location_NE). The named entities of a person type require
“not relevant.” In the lower right quadrant, “✓” means that
special attention a person can be mentioned not only using
the algorithm should be applied and “7” means that it should
pronouns but also using a position he/she holds (Posi-
not be applied.
tion_Held) and a professional name (Profession). Additional
The idea is that the pronoun-related coreferences should
information about a person could help resolving such
be solved first sequentially by checking the conditions C1, C2,
coreferences more precisely. As an example, source of such
and C3. Then a noun-related coreference resolution should
information could be a Database of Public Person, which
start by sequentially checking the conditions C4, C6, and C7.
includes Known_Person—a well-known person mentioned
as Person_NE in the text. The output of a coreference res-
olution algorithm is a Coreference—a relationship be- 3.4. Formal Description of Coreference Resolution Algorithms.
tween coreferents. For each coreference, its type (nominal First-order logic (FOL) formulas are employed to define
and pronominal), subtype (relative pronoun and noun the main conditions the algorithms should check when
repetition), position (points backward, forward, or irrelevant resolving coreferences. The concepts of the coreference
C1: Is a lexeme a pronoun? Yes No
C2: Does a specific rule exist for this Yes No na

pronoun?
C3: Was the pronoun resolved by specific Yes No na na

rules resolution?
Conditions
Answers
C4: Is a lexeme a noun? na na na Yes No
C5: Is the noun recognized as a named entity? na na na Yes No na
C6: Does the noun exist in profession na na na na Yes No na

classification?
C7: Does the noun exist in the knowledge
base of public persons? na na na na Yes No Yes No na
A1: Specific rules resolution

A2: General pronoun resolution
Algorithms
Decisions
A3: PRA resolution
A4: HHS resolution
A5: Feature resolution
Figure 3: A decision table for selection of the algorithm. na: not applicable; : the algorithm should be applied; : it should not be
applied.
resolution model (Figure 2) became the predicates or A condition for the existence of such reference formally
constants in the FOL formulas: the classes became the unary is defined as follows:
predicates of the same name as class; the associations be- For every sentence s of text t and for every “Relative” type
tween classes—the binary predicates of the same name as pronoun p, which is contained in the sentence s and has a start
association; the attributes of classes—the binary predicates position sp1, is of length ln1, follows comma c or follows
of the same name as attribute plus verb “has” at the be- prepositional lexeme l1, which follows comma c, and for every
ginning; and the literals of enumerations—constants. noun l2, which has a start position sp2, is of length ln2,
The algorithms follow the grammar rules of the Lithu- precedes comma c, is of the same gender g and of the same
anian language which are based on the analysis of mor- number n as the pronoun p, the only one coreference relation
phological features of lexemes and their order in the r, which is resolved in text t, is of “Pronominal” type, “Rel-
sentence and text. Examples of Lithuanian language sen- ative” subtype, “Backward” position and “Single” group be-
tences were translated into English as closely as possible. All tween the pronoun p and the noun n, its referent starts at
proper names were changed to generic abbreviations to position sp1 and has length ln1, and which fits only one
comply with GDPR. lexeme p and refers to only one mention m, which starts at
position sp2, has length ln2, and fits only one lexeme l2, exists
(Rule 1).
3.4.1. A1: Specific Rules Resolution. In some cases, there
exists a rather rigid structure for pronoun usage and it can be Rule 1: ∀t, s, p, l1, c, l2, g, n, sp1, sp2, ln2. [Text(t) ∧
easily defined by using specific rules, for example, Sentence(s) ∧ consists_of(t, s) ∧ Pronoun(p) ∧ con-
(i) [LT] Šiandien buvo atėje˛s vyras [noun], kuris tains(s, p) ∧ has_type(p, Relative) ∧ has_start_
[pronoun] skundėsi nugaros skausmu. position(p, sp1) ∧ has_length(p, ln1) ∧ Comma(c) ∧
(follows(p, c) ∨ (Lexeme(l1) ∧ has_pos(l1, Preposition)
[EN] A man [noun] who [pronoun] had a backache
∧ follows(l1, c) ∧ follows(p, l1)) ∧ Noun(l2) ∧ follows(l2,
came today.
c) ∧ has_gender (p, g) ∧ has_gender (l2, g) ∧ has_
(ii) [LT] Šiandien buvo atėje˛s vyras [noun], su [prep- number(p, n) ∧ has_number(l2, n) ∧ has_start_
osition] kuriuo [pronoun] aptarėme nugaros position(l2, sp2) ∧ has_length(l2, ln2) ⟶ ∃!r ∃!m.
skausma˛. [Coreference(r) ∧ resolved_in(r, t) ∧ has_type(r, Pro-
[EN] A man [noun] with [preposition] whom nominal) ∧ has_subtype(r, Relative) ∧ has_position(r,
[pronoun] we discussed a backache came today. Backward) ∧ has_ group(r, Single) ∧ has_start_
position(r, sp1) ∧ has_length(r, ln1) ∧ fits(r, p) ∧
Both examples are similar in their construction: [noun]
Mention(m) ∧ refers_to(r, m) ∧ has_start_position(m,
[comma] [optional preposition] [specific pronoun]. In both
sp2) ∧ has_length(m, ln2) ∧ fits(m, l2)]]
cases, pronoun “kuriuo” refers to the noun “vyras.” In the
first example, we do not have an optional preposition “su,” The relative pronoun might be plural and refer to
while we have it in the second one. multiple singular (or multiple plural) nouns:
(i) [LT] Komisija nerado panašumu˛ tarp Tomo [noun], resolution of such coreferences could be applied for other
Lino [noun], Petro [noun] ir [conjunction] Eglės languages as well. For example, Rule 1 could be successfully
[noun], kuriu˛ [pronoun] sužalojimai atrodė panaš�
us. applied for coreference resolution in Polish or Russian
[EN] The committee did not find Tom [noun], Linas languages. Let us take the same example of a sentence in
[noun], Peter [noun] and [conjunction] Eglė [noun], Polish and Russian:
who [pronoun] shared similar injuries. [PL] Dzisiaj przychodził me˛żczyzna [noun], który
In this case, a plural pronoun “kuriu˛” is referring to four [pronoun] skarżył sie˛ na ból pleców.
singular nouns that have different genders. The previous rule [RU] Sfгpeo> пrjypejm nuhyjoa [noun],
would not be able to solve such coreference. For this case, the lptpr9k [pronoun] hampcams> oa bpm: c sпjof.
construction would be: [noun] [comma] [noun] [comma]
We can see that a structure of the sentence (number and
[noun] [conjunction] [noun] [comma] [optional preposi-
order of lexemes) is similar, a pronoun goes after the comma
tion] [specific pronoun].
and it refers to a noun, and compatibility of morphological
For such case, a special condition must be defined:
features (gender, number) of noun and pronoun is retained.
For every sentence s in text t and for every “Relative” type
From the given example, we understand that the coreference
pronoun p of “Plural” number, which is contained in the
relation between pronoun and noun exists and conditions
sentence s and has a start position sp1, is of length ln1, follows
for such existence are the same as specified in Rule 1.
comma c1 or follows prepositional lexeme l, which follows
comma c1, and for every noun n1, which precedes comma c1,
has a start position sp2, is of length ln2, follows conjunction j, 3.4.2. A2: General Purpose Pronoun Resolution. This algo-
and for every noun n2, which precedes conjunction j, has a start rithm focuses on the cases where pronouns refer to nouns
position sp3, is of length ln3, and for every existing noun n3, (or noun phrases) that are recognized as named entities of
which follows comma c2, and for every existing noun n4, which “person” class by NER. The algorithm starts from the
precedes comma c2, has a start position sp4, is of length ln4, the identification of not demonstrative pronoun. In a given
only one coreference relation r, which is resolved in text t, is of example below, such a pronoun is in the second senten-
“Pronominal” type, “Relative” subtype, “Backward” position ce—“Jis” (“He”)
and “Multiple” group, its referent starts at position sp1 and has
length ln1, fits only one lexeme p, refers to only one mention (i) [LT] Jonas Jonaitis [person noun phrase] skambino ˛i
m1, which starts at position sp2, has length ln2, and fits noun registrat�
ura˛. Jis [pronoun] skundėsi galvos skausmu.
n1, refers to only one mention m2, which starts at position sp3, [EN] Jonas Jonaitis [person noun phrase] called
has length ln3, and fits only one noun n2, and refers at least to a reception. He [pronoun] complained about
one mention m3, which starts at position sp4, has length ln4, headache.
and fits noun n4, exists (Rule 2).
If the pronoun is in the relative clause, the algorithm
Rule 2: ∀t, s, p, l, c1, n1, sp1, ln1, sp2, ln2, j, n2, sp3, moves backwards analysing words going before the pro-
ln3.[Text(t) ∧ Sentence(s) ∧ consists_of(t, s) ∧ Pro- noun. In a given example, the pronoun is at the beginning of
noun(p) ∧ contains(s, p) ∧ has_number(p, Plural) ∧ has_ the sentence, so remaining parts of the sentence are not
type(p, Relative) ∧ has_start_position(p, sp1) ∧ has_ analysed, and the algorithm moves one sentence backwards.
length(p, ln1) ∧ Comma(c1) ∧ (follows(p, c1) ∨ (Lex- The conditions for the existence of such reference for-
eme(l) ∧ has_pos(l, Preposition) ∧ follows(p, l) ∧ fol- mally could be defined as three alternatives. The first one
lows(l, c1)) ∧ Noun(n1) ∧ follows(c1, n1) ∧ has_ describes conditions for reference existing in the same
start_position(n1, sp2) ∧ has_length(n1, ln2) ∧ Con- sentence s1 before pronoun p:
junction(j) ∧ follows(n1, j) ∧ Noun(n2) ∧ follows(j, n2) ∧ For each text’s t sentence s1 and pronoun p not of De-
has_start_position(n2, sp3) ∧ has_length(n2, ln3) ∧ monstrative type that is contained in sentence s1 and has
(∃n3, c2, n4, sp4, ln4.(Noun(n3) ∧ Comma(c2) ∧ gender g, number n, start position sp1 and length of ln1, and
Noun(n4) ∧ follows(n3, c2) ∧ follows(c2, n4) ∧ has_ named entity e1 that is in the same sentence s1, is expressed by
start_position(n4, sp4) ∧ has_length(n4, ln4)) ⟶ ∃!r ∃! lexeme l, and has gender g, number n, start position sp2 and is
m1 ∃!m2 ∃m3. [Coreference(r) ∧ resolved_in(r, t) ∧ of length ln2, and is before pronoun p (sp2 is lower than sp1),
has_type(r, Pronominal) ∧ has_subtype(r, Relative) ∧ but closer to pronoun p than possible named entities e2 and e3
has_position(r, Backward) ∧ has_ group(r, Multiple) (sp2 higher than sp3 and sp4), the only one coreference
∧ has_start_position(r, sp1) ∧ has_length(r, ln1) ∧ fits(r, relation r, which is resolved in text t, is of “Pronominal” type,
p) ∧ Mention(m1) ∧ refers_to(r, m1) ∧ has_ “Relative” subtype, “Backward” position and “Single” group
start_position(m1, sp2) ∧ has_length(m1, ln2) ∧ fits(m1, between the pronoun p and the named entity e1, its referent
n1) ∧ Mention(m2) ∧ refers_to(r, m2) ∧ has_start_ starts at position sp1 and has length ln1, and which fits only
position(m2, sp3) ∧ has_length(m2, ln3) ∧ fits(m2, n2) ∧ one pronoun p and refers to only one mention m, which starts
Mention(m3) ∧ refers_to(r, m3) ∧ has_start_ at position sp2, has length ln2, and fits only one named entity
position(m3, sp4) ∧ has_length(m3, ln4) ∧ fits(m3, n4)]] e1, exists (Rule 3).
Though examples illustrating the certain case of cor- Rule 3: ∀t, s1, p, l, e1, g, n, sp1, ln1, sp2, ln2.[Text(t) ∧
eference are given in Lithuanian and English only, rules for Sentence(s1) ∧ consists_of(t, s1) ∧ Pronoun(p) ∧
contains(s1, p) ∧ ¬has_type(p, Demonstrative) ∧ For each text’s t sentence s1, s2, s3, where s1 follows s2
has_gender(p, g) ∧ has_number(p, n) ∧ has_ and s2 follows s3, and pronoun p not of Demonstrative type
start_position(p, sp1) ∧ has_length(p, ln1) ∧ Person_ that is contained in sentence s1 and has gender g, number n,
NE(e1) ∧ includes(s1, e1) ∧ Lexeme(l) ∧ expressed_ start position sp1 and length of ln1, and named entity e1 that
by(e1, l) ∧ has_gender(e1, g) ∧ has_number(e1, n) ∧ is contained in sentence s3, is expressed by lexeme l, and has
has_start_position(e1, sp2) ∧ has_length(e1, ln2) ∧ gender g, number n, start position sp2 and is of length ln2,
sp2<sp1 ∧ ¬(∃e2, e3, sp3, sp4. (e1≠e2 ∧ e1≠e3 ∧ e2≠e3 and is closer to pronoun p than possible named entities e2 and
∧ Person_NE(e2) ∧ includes(s1, e2) ∧ has_gender(e2, e3 (sp2 higher than sp3 and sp4), the only one coreference
g) ∧ has_number(e2, n) ∧ has_start_position(e2, sp3) ∧ relation r, which is resolved in text t, is of “Pronominal” type,
Person_NE(p3) ∧ includes(s1, e3) ∧ has_gender(e3, g) “Relative” subtype, “Backward” position and “Single” group
∧ has_number(e3, n) ∧ has_start_position(e3, sp4) ∧ between the pronoun p and the named entity e1, its referent
sp2>sp3 ∧ sp4>sp2)) ⟶ ∃!r ∃!m. [Coreference(r) ∧ starts at position sp1 and has length ln1, and which fits only
resolved_in(r, t) ∧ has_type (r, Pronominal) ∧ has_ one pronoun p and refers to only one mention m, which starts
subtype (r, General) ∧ has_position(r, Backward) ∧ at position sp2, has length ln2, and fits only one named entity
has_group(r, Single) ∧ has_start_position(r, sp1) ∧ e1, exists (Rule 5).
has_length(r, ln1) ∧ fits(r, p) ∧ Mention(m) ∧ refer-
s_to(r, t) ∧ has_start_position(m, sp2) ∧ has_length(m, Rule 5: ∀t, s1, s2, s3, p, l, e1, g, n, sp1, ln1, sp2,
ln2) ∧ fits(m, e1) ∧ fits(m, l)]] ln2.[Text(t) ∧ Sentence(s1) ∧ Sentence(s2) ∧ Senten-
ce(s3) ∧ consists_of(t, s1) ∧ consists_of(t, s2) ∧ con-
The second alternative describes a case when a pronoun sists_of(t, s3) ∧ follows (s1, s2) ∧ follows (s2, s3) ∧
p refers to the named entity in the previous sentence s2: Pronoun(p) ∧ contains(s1, p) ∧ ¬has_type(p, De-
For each text’s t sentence s1, s2, where s1 follows s2, and monstrative) ∧ has_gender(p, g) ∧ has_number(p, n) ∧
pronoun p not of Demonstrative type that is contained in has_start_position(p, sp1) ∧ has_length(p, ln1) ∧
sentence s1 and has gender g, number n, start position sp1 and Person_NE(e1) ∧ Lexeme(l) ∧ expressed_by(e1, l) ∧
length of ln1, and named entity e1 that is contained in sentence includes(s3, e1) ∧ has_gender(e1, g) ∧ has_number(e1,
s2, is expressed by lexeme l, and has gender g, number n, start n) ∧ has_start_position(e1, sp2) ∧ has_length(e1, ln2) ∧
position sp2 and is of length ln2, and is closer to pronoun p ¬(∃e2, e3, sp3, sp4. (e1≠e2 ∧ e1≠e3 ∧ e2≠e3 ∧ Per-
than possible named entities e2 and e3 (sp2 higher than sp3 son_NE(e2) ∧ includes(s3, e2) ∧ has_gender(e2, g) ∧
and sp4), the only one coreference relation r, which is resolved has_number(e2, n) ∧ has_start_position(e2, sp3) ∧
in text t, is of “Pronominal” type, “Relative” subtype, “Back- Person_NE(e3) ∧ includes(s3, e3) ∧ has_gender(e3, g)
ward” position and “Single” group between the pronoun p and ∧ has_number(e3, n) ∧ has_start_position(e3, sp4) ∧
the named entity e1, its referent starts at position sp1 and has sp2>sp3 ∧ sp4>sp2)) ⟶ ∃!r ∃!m. [Coreference(r) ∧
length ln1, and which fits only one pronoun p and refers to only resolved_in(r, t) ∧ has_type (r, Pronominal) ∧ has_
one mention m, which starts at position sp2, has length ln2, subtype (r, General) ∧ has_position(r, Backward) ∧
and fits only one named entity e1, exists (Rule 4). has_group(r, Single) ∧ has_start_position(r, sp1) ∧
has_length(r, ln1) ∧ fits(r, p) ∧ Mention(m) ∧ refer-
Rule 4: ∀t, s1, s2, p, l, e1, g, n, sp1, ln1, sp2, ln2.[Text(t) ∧ s_to(r, t) ∧ has_start_position(m, sp2) ∧ has_length(m,
Sentence(s1) ∧ Sentence(s2) ∧ consists_of(t, s1) ∧ con- ln2) ∧ fits(m, e1) ∧ fits(m, l)]]
sists_of(t, s2) ∧ follows (s1, s2) ∧ Pronoun(p) ∧ con-
tains(s1, p) ∧ ¬has_type(p, Demonstrative) ∧ has_ Another example presents a case when a coreferent of
gender(p, g) ∧ has_number(p, n) ∧ has_start_position(p, the pronoun “man” (in English, “for me”) is in the following
sp1) ∧ has_length(p, ln1) ∧ Person_NE(e1) ∧ inclu- sentence:
des(s2, e1) ∧ Lexeme(l) ∧ expressed_by(e1, l) ∧ has_ (i) [LT] Pastebėtina, kad ligoginėse apsilankė 10 mln.
gender(e1, g) ∧ has_number(e1, n) ∧ has_start_ pacientu˛, nepaisant to, kad 2016 m. ju˛ buvo 4%
position(e1, sp2) ∧ has_length(e1, ln2) ∧ ¬(∃e2, e3, sp3, mažiau (apsilankė beveik 9,5 mln.). Tai labiausiai
sp4. (e1≠e2 ∧ e1≠e3 ∧ e2≠e3 ∧ Person_NE(e2) ∧ lėmė skaitmeniniu˛ paslaugu˛ padidinimas: “Kiek man
includes(s2, e2) ∧ has_gender(e2, g) ∧ has_number(e2, [pronoun] teko analizuoti, padidinus skaitmenines
n) ∧ has_start_position(e2, sp3) ∧ Person_NE(p3) ∧ paslaugas tik nedidelė dalis Lietuvos [location noun]
includes(s2, e3) ∧ has_gender(e3, g) ∧ has_number(e3, ligoniniu˛ sumažino etatu˛ ar atleido darbuotojus, o tai
n) ∧ has_start_position(e3, sp4) ∧ sp2>sp3 ∧ sp4>sp2)) lėmė nemaža˛ papildoma˛ indėli˛ ˛i paslaugos kokybe˛”
⟶ ∃!r ∃!m. [Coreference(r) ∧ resolved_in(r, t) ∧ has_ teigė J. Jonaitis [person noun phrase].
type (r, Pronominal) ∧ has_subtype (r, General) ∧
[EN] It is noteworthy that the total hospital patient
has_position(r, Backward) ∧ has_group(r, Single) ∧ has_
count has reached 10 million, even though in 2016 the
start_position(r, sp1) ∧ has_length(r, ln1) ∧ fits(r, p) ∧
number was less than 4% (around 9.5 million pa-
Mention(m) ∧ refers_to(r, t) ∧ has_start_position(m,
tients). This was influenced by the digitization of
sp2) ∧ has_length(m, ln2) ∧ fits(m, e1) ∧ fits(m, l)]]
e-health services. “As far as I [pronoun] had analysed,
The third alternative describes a case when a pronoun p only a small part of Lithuanian [location noun]
in the sentence s1 refers to the named entity in the sentence hospitals have reduced their posts or dismissed
s3, preceding sentences s2 and s1: employees, but digitization of services has led to a
significant additional contribution to the quality of and sp4 ), the only coreference relation r, which is resolved in
service,” said J. Jonaitis [person noun phrase]. text t, is of “Pronominal” type, “Relative” subtype, “Backward”
position and “Single” group between the pronoun p and the
If the algorithm does not find any named entities moving
named entity e1, its referent starts at position sp1 and has
backwards, it moves back to pronoun and proceeds forward.
length ln1, and which fits only one pronoun p and refers to only
The algorithm continues moving forward until it locates “J.
one mention m, which starts at position sp2, has length ln2,
Jonaitis” entity, which is recognized as a person. Since the
and fits only one named entity e1, exists (Rule 7).
gender of the pronoun “man” is ambiguous (it can refer to
both female and male persons), only their grammatical Rule 7: ∀t, s1, s2, p, l, e1, g, n, sp1, ln1, sp2, ln2.[Text(t) ∧
numbers are compared. Both are singular; therefore, the Sentence(s1) ∧ Sentence(s2) ∧ consists_of(t, s1) ∧
algorithm picks “J. Jonaitis” as a postcedent of the corefering consists_of(t, s2) ∧ follows (s2, s1) ∧ Pronoun(p) ∧
object “man.” Conditions for the existence of such reference contains(s1, p) ∧ ¬has_type(p, Demonstrative) ∧ has_
formally could be defined as two alternatives. The first one gender(p, g) ∧ has_number(p, n) ∧ has_start_
describes the conditions for reference existing in the same position(p, sp1) ∧ has_length(p, ln1) ∧ Person_NE(e1)
sentence s1 after pronoun was mentioned: ∧ includes(s2, e1) ∧ Lexeme(l) ∧ expressed_by(e1, l) ∧
For each text’s t sentence s1 and pronoun p not of De- has_gender(e1, g) ∧ has_number(e1, n) ∧ has_start_
monstrative type that is contained in sentence s1 and has position(e1, sp2) ∧ has_length(e1, ln2) ∧ ¬(∃e2, e3, sp3,
gender g, number n, start position sp1 and length of ln1, and sp4. (e1≠e2 ∧ e1≠e3 ∧ e2≠e3 ∧ Person_NE(e2) ∧
named entity e1 that is in the same sentence s1, is expressed by includes(s2, e2) ∧ has_start_position(e2, sp3) ∧ Per-
lexeme l, and has gender g, number n, start position sp2 and is son_NE(e3) ∧ includes(s2, e3) ∧ has_start_position(e3,
of length ln2, and is after pronoun p (sp2 is higher than sp1), sp4) ∧ sp2<sp3 ∧ sp4<sp2)) ⟶ ∃!r ∃!m. [Corefer-
but closer to pronoun p than possible named entities e2 and e3 ence(r) ∧ resolved_in(r, t) ∧ has_type (r, Pronominal) ∧
(sp2 higher than sp3 and sp4), the only one coreference has_subtype (r, General) ∧ has_position(r, Forward) ∧
relation r, which is resolved in text t, is of “Pronominal” type, has_group(r, Single) ∧ has_start_position(r, sp1) ∧
“Relative” subtype, “Backward” position and “Single” group has_length(r, ln1) ∧ fits(r, p) ∧ Mention(m) ∧ refer-
between the pronoun p and the named entity e1, its referent s_to(r, t) ∧ has_start_position(m, sp2) ∧ has_length(m,
starts at position sp1 and has length ln1, and which fits only ln2) ∧ fits(m, e1) ∧ fits(m, l)]]
one pronoun p and refers to only one mention m, which starts
at position sp2, has length ln2, and fits only one named entity The algorithm ignores demonstrative pronouns because
e1, exists (Rule 6). they are often used to refer to entities that are not present in
the written text. Such pronouns do not carry any additional
Rule 6: ∀t, s1, p, l, e1, g, n, sp1, ln1, sp2, ln2.[Text(t) ∧ semantic information and do not refer to any noun phrase.
Sentence(s1) ∧ consists_of(t, s1) ∧ Pronoun(p) ∧ They are used mostly for syntactic reasons and due to that
contains(s1, p) ∧ ¬has_type(p, Demonstrative) ∧ has_ are usually ignored in coreference resolution.
gender(p, g) ∧ has_number(p, n) ∧ has_start_
position(p, sp1) ∧ has_length(p, ln1) ∧ Person_NE(e1)
∧ includes(s1, e1) ∧ Lexeme(l) ∧ expressed_by(e1, l) ∧ 3.4.3. A3: PRA Resolution. This algorithm is based on exact
has_gender(e1, g) ∧ has_number(e1, n) ∧ has_ (or partial) string matches and several rules for acronyms.
start_position(e1, sp2) ∧ has_length(e1, ln2) ∧ sp1<sp2 Once a first named entity that can be matched with an initial
∧ ¬(∃e2, e3, sp3, sp4. (e1≠e2 ∧ e1≠e3 ∧ e2≠e3 ∧ named entity is found, then the algorithm stops to keep
Person_NE(e2) ∧ includes(s1, e2) ∧ has_start_ annotations simple: B ⟶ A, C ⟶ B and D ⟶ C. This
position(e2, sp3) ∧ Person_NE(e3) ∧ includes(s1, e3) ∧ allows the formation of the coreference chains linking all
has_start_position(e3, sp4) ∧ sp2<sp3 ∧ sp4<sp2)) ⟶ mentions of the same entity in a text that can be later reused
∃!r ∃!m. [Coreference(r) ∧ resolved_in(r, t) ∧ has_type for semantic analysis, for example,
(r, Pronominal) ∧ has_subtype (r, General) ∧ has_
position(r, Forward) ∧ has_group(r, Single) ∧ has_ (i) [LT] Tomaitis [named entity] pateko ˛i avarija˛. Po
start_position(r, sp1) ∧ has_length(r, ln1) ∧ fits(r, p) ∧ pietu˛ Tomaiti˛ [named entity] išvežė ˛i operacine˛.
Mention(m) ∧ refers_to(r, t) ∧ has_start_position(m, [EN] Tomaitis [named entity] got into a car accident.
sp2) ∧ has_length(m, ln2) ∧ fits(m, e1) ∧ fits(m, l)]] In the afternoon, Tomaitis [named entity] has been
taken to a surgery room.
The second alternative describes the case when the
pronoun p refers to the named entity in the following In this example, two mentions of the same entity are
sentence s4: made: “Tomaitis” and “Tomaiti˛.” They are of different cases,
For each text’s t sentence s1, s4, where s4 follows s1, and but their lemmas are identical. A condition for the existence
pronoun p not of Demonstrative type that is contained in of such reference formally is defined as follows:
sentence s1 and has gender g, number n, start position sp1 and For each text’s t sentence s1 that includes named entity
length of ln1, and named entity e1 that is contained in sentence e1, that has start position sp1 and is of length ln1, which is
s2, is expressed by lexeme l, and has gender g, number n, start expressed by lexeme l1 that has lemma l and for each same
position sp2 and is of length ln2, and is closer to pronoun p text’s t sentence s2 that includes named entity e2, that has a
than possible named entities e2 and e3 ( sp2 higher than sp3 start position sp1 and is of length ln1, which is expressed by
lexeme l2 that has lemma l, the only one coreference relation Rule 9: ∀t, s1, s2, n1, n2, sp1, ln1, sp2, ln2, v1, v2, p1,
r, which is resolved in text t, is of “Nominal” type, “Repe- p2.[Text(t) ∧ Sentence(s1) ∧ Sentence(s2) ∧ consists_
tition” subtype, “Irrelevant” position and “Single” group of(t, s1) ∧ consists_of(t, s2) ∧ Noun(n1) ∧ contains(s,
between the noun n1 and the noun n2, its referent starts at n1) ∧ has_start_position(n1, sp1) ∧ has_length(n1, ln1)
position sp1 and has length ln1, and which fits only one ∧ Noun(n2) ∧ contains(s2, n2) ∧ has_start_position(n2,
noun n1 and refers to only one mention m, which starts at sp2) ∧ has_length(n2, ln2) ∧ n1≠n2 ∧ Profession(p1) ∧
position sp2, has length ln2, and fits only one noun n2, exists Profession(p2) ∧ p1≠p2 ∧ Profession_Name(v1) ∧
(Rule 8). Profession_Name(v2) ∧ express(n1, v1) ∧ express(n2,
v2) ∧ describes(v1, p1) ∧ describes(v2, p2) ∧ (broad-
Rule 8: ∀t, s1, s2, e1, e2, sp1, ln1, sp2, ln2.[Text(t) ∧
ens(p2, p1) ∨ broadens(p1, p2)) ∧ has_gender(n1, g) ∧
Sentence(s1) ∧ Sentence(s2) ∧ consists_of(t, s1) ∧
has_gender(n2, g) ∧ has_number(n1, n) ∧ has_num-
consists_of(t, s2) ∧ Named_Entity(e1) ∧ includes(s1,
ber(n2, n) ⟶ ∃!r ∃!m. [Coreference(r) ∧ resolved_in(r,
e1) ∧ has_start_position(e1, sp1) ∧ has_length(e1, ln1)
t) ∧ has_type (r, Nominal) ∧ has_subtype (r, Hyper-
∧ Named_Entity(e2) ∧ includes(s2, e2) ∧ has_-
nym_Hyponym) ∧ has_position(r, Irrelevant) ∧ has_
start_position(e2, sp2) ∧ has_length(e2, ln2) ∧ e1≠e2 ∧
group(r, Single) ∧ has_start_position(r, sp1) ∧ has_
(∃l1 ∃l2 ∃l.(Lexeme(l1) ∧ Lexeme(l2) ∧ expressed_-
length(r, ln1) ∧ fits(r, n1) ∧ Mention(m) ∧ refers_to(r, t)
by(e1, l1) ∧ expressed_by(e2, l2) ∧ has_lemma(l1, l) ∧
∧ has_start_position(m, sp2) ∧ has_length(m, ln2) ∧
has_lemma(l2, l)) ⟶ ∃!r ∃!m. [Coreference(r) ∧
fits(m, n2)]]
resolved_in(r, t) ∧ has_type (r, Nominal) ∧ has_sub-
type (r, Repetition) ∧ has_position(r, Irrelevant) ∧ An example of synonym is given as follows:
has_group(r, Single) ∧ has_start_position(r, sp1) ∧
(i) [LT] J. Jonaitis buvo operacijos vadovu [noun re-
has_length(r, ln1) ∧ fits(r, l1) ∧ fits(r, e1) ∧ Mention(m)
ferring to profession]. Deja, vyr. chirurgo [noun
∧ refers_to(r, t) ∧ has_start_position(m, sp2) ∧ has_-
referring to profession] vykdoma operacija buvo
length(m, ln2) ∧ fits(m, l2) ∧ fits(m, e2)]]
nesėkminga.
Acronym rules vary depending on the type of named [EN] From now J. Jonaitis was the head surgeon
entity (currently persons, locations, and organizations are [noun referring to profession] of operation. Un-
covered). fortunately, the last operation of chief surgeon [noun
referring to profession] was unsuccessful.
3.4.4. A4: HHS Resolution. This algorithm is based on Both “head surgeon” and “chief surgeon” are synony-
profession classification. It attempts to resolve the use of mous; therefore, the condition for the existence of such
synonyms and hypernyms/hyponyms. reference formally could be defined as follows:
For each text’s t sentence s1 that has a profession’s p name
(i) [LT] Gydytojai [noun referring to profession] v1, which is expressed by noun n1 that has gender g, number m,
skundžiasi dideliu darbo kr� uviu. Chirurgu˛ [noun start position sp1 and is of length ln1, and for each same text’s t
referring to profession] darbo kr�
uvis pats didžiausias. sentence s2 that has same profession’s p name v2 expressed by
[EN] Doctors [noun referring to profession] com- noun n2 that has gender g, number m, start position sp2 and is
plain about heavy workload. Surgeons’ [noun re- of length ln2, the only one coreference relation r, which is
ferring to profession] workload is the greatest. resolved in text t, is of “Nominal” type, “Synonym” subtype,
“Irrelevant” position and “Single” group between the noun n1
The algorithm determines that “Doctor” in professions
and the noun n2, its referent starts at position sp1 and has
classification is a hyponym of “Surgeon,” they also agree in
length ln1, and which fits only one noun n1 and refers to only
gender and number; therefore, the algorithm adds their pair
one mention m, which starts at position sp2, has length ln2, and
to annotations. Conditions for the existence of such refer-
fits only one noun n2, exists (Rule 10).
ence formally are defined as follows:
For each text’s t sentence s1 that has profession p1, which is Rule 10: ∀t, s1, s2, n1, n2, sp1, ln1, sp2, ln2, v1, v2, p, g,
either broader or narrower than profession p2, name v1 n.[Text(t) ∧ Sentence(s1) ∧ Sentence(s2) ∧ consists_
expressing noun n1, which has gender g, number m, start of(t, s1) ∧ consists_of(t, s2) ∧ Noun(n1) ∧ contains(s1,
position sp1 and is of length ln1, and for each same text’s t n1) ∧ has_start_position(n1, sp1) ∧ has_length(n1, ln1)
sentence s2 that has profession p2, which is either broader or ∧ Noun(n2) ∧ contains(s2, n2) ∧ has_start_
narrower than profession p1, name v2 expressing noun n2, position(n2, sp2) ∧ has_length(n2, ln2) ∧ n1≠n2 ∧
which has gender g, number m, start position sp2 and is of Profession_name(v1) ∧ Profession_name(v2) ∧ Pro-
length ln2, the only one coreference relation r, which is resolved fession(p) ∧ express(n1, v1) ∧ express(n2, v2) ∧
in text t, is of “Nominal” type, “Hypernym_hyponym” subtype, describes(v1, p) ∧ describes(v2, p) ∧ has_gender(n1, g)
“Irrelevant” position and “Single” group between the noun n1 ∧ has_gender(n2, g) ∧ has_number(n1, n) ∧ has_
and the noun n2, its referent starts at position sp1 and has number(n2, n) ∧ n1≠n2 ⟶ ∃!r ∃!m. [Coreference(r) ∧
length ln1, and which fits only one noun n1 and refers to only resolved_in(r, t) ∧ has_type (r, Nominal) ∧ has_sub-
one mention m, which starts at position sp2, has length ln2, type(r, Synonym) ∧ has_position(r, Irrelevant) ∧
and fits only one noun n2, exists (Rule 9). has_group(r, Single) ∧ has_start_position(r, sp1) ∧
has_length(r, ln1) ∧ fits(r, n1) ∧ Mention(m) ∧ refers_ group(r, Single) ∧ has_start_position(r, sp2) ∧ has_-
to(r, t) ∧ has_start_position(m, sp2) ∧ has_length(m, length(r, ln2) ∧ fits(r, e) ∧ Mention(m) ∧ refers_to(r, t)
ln2) ∧ fits(m, n2)]] ∧ has_start_position(m, sp1) ∧ has_length(m, ln1) ∧
fits(m, n)]]
In this case, it is also relevant to track if coreference is
3.4.5. A5: Feature Resolution. This algorithm at the time pointing backward or forward. We can rewrite the same
attempts to resolve only those cases when a person is being
example and switch a known person with his positions:
referred to by his public post (feature) that he holds, other
types of features are not currently resolved, for example, (i) [LT] Koks yra mano šeimos gydytojas [noun re-
ferring to held position]? T. Tomaitis [person noun
(i) [LT] Ka˛ rekomenduoja S. Suskelis [person noun phrase] yra labai patyre˛s j�
usu˛ šeimos gydytojas.
phrase]? Aptarkime kardiologo [noun referring to
[EN] Who is my family doctor [noun referring to
held position] si�
uloma˛ gydymo plana˛.
held position]? T. Tomaitis [person noun phrase] is
[EN] What does S. Suskelis [person noun phrase] very experienced doctor assigned to you.
recommend? Let’s discuss treatment plan proposed
by the cardiologist [noun referring to held position]. As a result, sp1 is higher than sp2 and coreference has
Here a noun “cardiologist” is selected, the algorithm different position constant values:
moves backwards till it reaches “S. Suskelis” and For each text’s t sentence s1 that has known person k,
checks the knowledge base if at the time of the who during publication date d had certain position h
publication of the medical record he has held the (publication date d is same or later than position h start date
position of the cardiologist.. Since “he holds it” the fd and same or earlier than position h end date td), mention
algorithm checks if “S. Suskelis” and “cardiologist” as named entity e, that has a start position sp1 and is of
agree in gender and number. They agree, and their length ln1, and for each same text’s t sentence s2 mentioned
pair is added to annotation as a feature reference. A noun n, that has a start position sp2 and is length ln2, which
condition for the existence of such reference formally is mentioned after named entity e (noun n has a lower start
is defined as follows: position sp2 than named entity’s sp1), whose lemma l
matches with position’s h lemma l, number is Singular and
For each text’s t sentence s1 that has known person k, gender g matches known person’s k gender g, the only one
who during publication date d had certain position h coreference relation r, which is resolved in text t, is of
(publication date d is same or later than position h start date “Nominal” type, “Feature” subtype, “Forward” position and
fd and same or earlier than position h end date td), mention “Single” group between the noun n and the named entity e,
as named entity e, that has a start position sp1 and is of its referent starts at position sp2 and has length ln2, and
length ln1, and for each same text’s t sentence s2 mentioned which fits only one noun n and refers to only one mention m,
noun n, that has a start position sp2 and is length ln2, which which starts at position sp1, has length ln1, and fits only one
is mentioned after named entity e (noun n has a higher start named entity e, exists (Rule 12).
position sp2 than named entity’s sp1), whose lemma l
matches with position’s h lemma l, number is Singular and Rule 12: ∀t, s1, s2, n, k, h, l, d, fd, td, g, sp1, sp2, ln1,
gender g matches known person’s k gender g, the only one ln2.[Text(t) ∧ Sentence(s1) ∧ Sentence(s2) ∧ con-
coreference relation r, which is resolved in text t, is of sists_of(t, s1) ∧ consists_of(t, s2) ∧ Person_NE(e) ∧
“Nominal” type, “Feature” subtype, “Backward” position includes(s1, e) ∧ Noun(n) ∧ contains(s2, n) ∧
and “Single” group between the noun n and the named entity Known_Person(k) ∧ mentioned_as(k, e) ∧ Position_
e, its referent starts at position sp2 and has length ln2, and held(h) ∧ holds(k, h) ∧ has_lemma(h, l) ∧ has_lem-
which fits only one noun n and refers to only one mention m, ma(n, l) ∧ has_publication_date(t, d) ∧ has_from_
which starts at position sp1, has length ln1, and fits only one date(h, fd) ∧ has_to_date(h, td) ∧ fd≤d ∧ td≥d ∧
named entity e, exists (Rule 11). has_gender(k, g) ∧ has_gender(n, g) ∧ has_number(n,
Singular) ∧ has_start_position(e, sp1) ∧ has_
Rule 11: ∀t, s1, s2, n, k, h, l, d, fd, td, g, sp1, sp2, ln1, start_position(n, sp2) ∧ sp1>sp2 ∧ has_length(e, ln1) ∧
ln2.[Text(t) ∧ Sentence(s1) ∧ Sentence(s2) ∧ con- has_length(n, ln2) ⟶ ∃!r ∃!m. [Coreference(r) ∧
sists_of(t, s1) ∧ consists_of(t, s2) ∧ Person_NE(e) ∧ resolved_in(r, t) ∧ has_type (r, Nominal) ∧ has_sub-
includes(s1, e) ∧ Noun(n) ∧ contains(s2, n) ∧ type (r, Feature) ∧ has_position(r, Forward) ∧ has_
Known_Person(k) ∧ mentioned_as(k, e) ∧ Position_ group(r, Single) ∧ has_start_position(r, sp2) ∧ has_
held(h) ∧ holds(k, h) ∧ has_lemma(h, l) ∧ has_lem- length(r, ln2) ∧ fits(r, e) ∧ Mention(m) ∧ refers_to(r, t)
ma(n, l) ∧ has_publication_date(t, d) ∧ has_from_ ∧ has_start_position(m, sp1) ∧ has_length(m, ln1) ∧
date(h, fd) ∧ has_to_date(h, td) ∧ fd≤d ∧ td≥d ∧ fits(m, n)]]
has_gender(k, g) ∧ has_gender(n, g) ∧ has_number(n,
Singular) ∧ has_start_position(e, sp1) ∧ has_ 4. Results and Evaluation
start_position(n, sp2) ∧ sp1<sp2 ∧ has_length(e, ln1) ∧
has_length(n, ln2) ⟶ ∃!r ∃!m. [Coreference(r) ∧ The coreference resolution algorithms and rules presented in
resolved_in(r, t) ∧ has_type (r, Nominal) ∧ has_sub- Section 3 were implemented as a separate component and
type (r, Feature) ∧ has_position(r, Backward) ∧ has_ integrated into the semantic search framework NLP pipeline
Table 3: Experiment results.

F (number of C (number of
T (actual number
solvable expressions solvable expressions
Algorithm of solvable R (recall) P (precision) F1
resolved by correctly resolved
expressions)
the algorithm) by the algorithm)
General purpose (GP) 648 248 202 0.311 0.814 0.451
GP + specific usage rules (SUR) 648 277 223 0.344 0.805 0.482
GP + SUR + PRA 648 330 267 0.412 0.809 0.546
GP + SUR + PRA + HHS 648 343 274 0.422 0.799 0.553
GP + SUR + PRA + HHS + feature 648 371 289 0.446 0.779 0.567
(Figure 1) because it requires lexical, morphological, and NE written about things that happened in the past. There are no
annotations of the text should be analysed. Solutions for tools, which can identify the timeframe of a certain part of
other languages should not follow the same NLP pipeline the text.
architecture. But a supply of coreference resolution com-
ponent with lexical, morphological, and NE information of 5. Conclusion
the text must be ensured.
Coreference resolution for Lithuanian was implemented Medical entity recognition and coreferencing are difficult
using Java programming language and JSON data format for tasks in Lithuanian natural language processing (NLP). We
annotation storage. But the proposed approach is not proposed the coreference resolution approach for the
technology dependent, and for other languages, it can be Lithuanian language. The coreference resolution algorithm
implemented on any other platform. depends on morphological and named entity recognition
The evaluation was performed by analysing 100 articles (NER) annotations and preexisting databases. Due to the
that have been preannotated and are available in our proposed approach being detached from specific imple-
Lithuanian Language Coreference Corpus [55], in addition mentation and rules being formalized, it would not be
to the transcribed records of medical reception, which we difficult to adapt it for grammatically similar languages. Our
cannot disclose due to the privacy requirements. novelty is the ability to process coreferences with minimal
For evaluation, we used precision, recall, and F1 metrics. linguistic resources, which are very important to consider in
Recall R is the ratio of correctly resolved anaphoric ex- linguistic applications for under-resourced and endangered
pressions C to the total number of anaphoric expressions T. languages. While the proposed method provides encour-
Precision P is the ratio of correctly resolved anaphoric aging results, when analysing transcribed medical records
expressions C to the number of resolved anaphoric ex- and other corpora, and they are comparable to the results
pressions F. F1 is a harmonic mean of P and R: achieved by other authors applying different resolution
C approaches on other languages, it has certain limitations: it is
R� , domain specific and is able to resolve only a subset of
T
coreference types, while the relatively small dataset was used
C for experiments. Nevertheless, we hope that our method can
P� , (1) contribute to the sustainable development of the NLP-
F
powered online healthcare services in Lithuania.
R∗P
F1 � 2 ∗ .
R+P Data Availability
Five experiments were performed with different com- The dataset used in this research is available upon request.
binations of coreferencing algorithms presented in Section 3.
The results of the experiments are presented in Table 3.
Conflicts of Interest
Note the following threats to validity of our results:
(i) The database of public persons must be constantly The authors declare that they have no conflicts of interest.
updated as new information becomes available.
Otherwise, recall will get noticeably lower when References
annotating newer texts.
[1] A. Pomares-Quimbaya, R. A. Gonzalez, S. Quintero et al., “A
(ii) In the case where plural pronouns and nouns are review of existing applications and techniques for narrative
used, they are difficult to be identified because of text analysis in electronic medical records,” in Artificial In-
many variations possible that often ignore gram- telligence, pp. 1620–1635, Optimizing Medicine Residency
matical compatibility rules. Training Programs, Hershey, PA, USA, 2017.
[2] B. L. Cook, A. M. Progovac, P. Chen, B. Mullin, S. Hou, and
Linking the named entity to the position held taking into E. Baca-Garcia, “Novel use of natural language processing
account the date of the publication of the text is limited (NLP) to predict suicidal ideation and psychiatric symptoms
considering that the text might be published today but in a text-based mental health intervention in madrid,”
Computational and Mathematical Methods in Medicine, [18] S. Doan, M. Conway, T. M. Phuong, and L. Ohno-Machado,
vol. 2016, p. 8, 2016. “Natural language processing in biomedicine: a unified system
[3] J. Ginsberg, M. H. Mohebbi, R. S. Patel, L. Brammer, architecture overview,” in Methods in Molecular Biology,
M. S. Smolinski, and L. Brilliant, “Detecting influenza epi- pp. 275–294, Springer, New York, NY, USA, 2014.
demics using search engine query data,” Nature, vol. 457, [19] W. Sun, Z. Cai, Y. Li, F. Liu, S. Fang, and G. Wang, “Data
no. 7232, pp. 1012–1014, 2009. processing and text mining technologies on electronic
[4] A. Nikfarjam, A. Sarker, K. O’Connor, R. Ginn, and medical records: a review,” Journal of Healthcare Engineering,
G. Gonzalez, “Pharmacovigilance from social media: mining vol. 2018, Article ID 4302425, 9 pages, 2018.
adverse drug reaction mentions using sequence labeling with [20] Z. He, C. Tao, J. Bian, M. Dumontier, and W. R. Hogan,
word embedding cluster features,” Journal of the American “Semantics-powered healthcare engineering and data ana-
Medical Informatics Association, vol. 22, pp. 671–681, 2015. lytics,” Journal of Healthcare Engineering, vol. 2017, Article ID
[5] H. Yang and H. Gao, “Toward sustainable virtualized 7983473, 3 pages, 2017.
healthcare: extracting medical entities from Chinese online [21] O. Bodenreider, “The unified Medical Language system
health consultations using deep neural networks,” Sustain- (UMLS): integrating biomedical terminology,” Nucleic Acids
ability, vol. 10, no. 9, p. 3292, 2018. Research, vol. 32, no. 90001, pp. D267–D270, 2004.
[6] D. Ferrucci, E. Brown, J. Chu-Carroll et al., “Building Watson: [22] G. Vanagas and R. Klimaviči� utė-Gudauskiene, “Factors af-
an overview of the DeepQA project,” AI magazine, vol. 31, fecting electronic health information needs in primary care
no. 3, pp. 59–79, 2010. patients,” Telemedicine and e-Health, vol. 18, no. 10,
[7] A. Utka, D. Amilevicius, T. Krilavicius, and D. Vitkute- pp. 724–728, 2012.
Adzgauskiene, “Overview of the development of language [23] T. Vileiniškis, A. Šukys, and R. Butkienė, “Searching the web
resources and technologies in Lithuania (2012–2015),” by meaning: a case study of Lithuanian news websites,” in
Frontiers in Artificial Intelligence and Applications, vol. 289, Proceedings of the 7th International Joint Conference on
pp. 12–19, 2016. Knowledge Discovery, Knowledge Engineering and Knowledge
[8] M. Kholghi, L. Sitbon, G. Zuccon, and A. Nguyen, “Active Management: IC3K 2015, vol. 631, pp. 47–64, Springer, Porto,
learning: a step towards automating medical concept ex- Portugal, November 2016.
traction,” Journal of the American Medical Informatics As- [24] O. Uzuner, A. Bodnari, S. Shen, T. Forbush, J. Pestian, and
sociation, vol. 23, no. 2, pp. 289–296, 2016. B. R. South, “Evaluating the state of the art in coreference
[9] E. Choi, M. T. Bahadori, E. Searles et al., “Multi-layer rep- resolution for electronic medical records,” Journal of the
resentation learning for medical concepts,” in Proceedings of American Medical Informatics Association, vol. 19, no. 5,
the 22nd ACM SIGKDD International Conference on pp. 786–791, 2012.
Knowledge Discovery and Data Mining - KDD’16, pp. 1495– [25] H. Lee, M. Surdeanu, and D. Jurafsky, “A scaffolding approach
1504, San Francisco, CA, USA, August 2016. to coreference resolution integrating statistical and rule-based
[10] Y.-M. Kim and D. Delen, “Medical informatics research trend models,” Natural Language Engineering, vol. 23, no. 5,
analysis: a text mining approach,” Health Informatics Journal, pp. 733–762, 2017.
vol. 24, no. 4, pp. 432–452, 2018. [26] S. Liu, B. Tang, Q. Chen, and X. Wang, “Effects of se-
[11] M. Simmons, A. Singhal, and Z. Lu, “Text mining for pre- mantic features on machine learning-based drug name
cision medicine: bringing structure to EHRs and biomedical recognition systems: word embeddings vs. Manually
literature to understand genes and health,” Advances in constructed dictionaries,” Information, vol. 6, no. 4,
Experimental Medicine and Biology, vol. 939, pp. 139–166, pp. 848–865, 2015.
2016. [27] C. Li, D. Konomis, G. Neubig, P. Xie, C. Cheng, and E. Xing,
[12] S. Liu, B. Tang, Q. Chen, X. Wang, and X. Fan, “Feature “Convolutional neural networks for medical diagnosis from
engineering for drug name recognition in biomedical texts: admission notes,” 2017, https://arxiv.org/abs/1712.02768v1.
feature conjunction and feature selection,” Computational [28] H. Ranjan, S. Agarwal, A. Prakash, and S. K. Saha, “Automatic
and Mathematical Methods in Medicine, vol. 2015, Article ID labelling of important terms and phrases from medical dis-
913489, 9 pages, 2015. cussions,” in Proceedings of the 2017 Conference on In-
[13] D. Zeng, C. Sun, L. Lin, and B. Liu, “LSTM-CRF for drug- formation and Communication Technology (CICT), pp. 1–5,
named entity recognition,” Entropy, vol. 19, no. 6, p. 283, 2017. Gwalior, India, November 2017.
[14] B. Tang, H. Cao, X. Wang, Q. Chen, and H. Xu, “Evaluating [29] E. Chiaramello, F. Pinciroli, A. Bonalumi, A. Caroli, and
word representation features in biomedical named entity G. Tognola, “Use of “off-the-shelf” information extraction
recognition tasks,” BioMed Research International, vol. 2014, algorithms in clinical informatics: a feasibility study of
Article ID 240403, 6 pages, 2014. MetaMap annotation of Italian medical notes,” Journal of
[15] H.-J. Dai, S. Syed-Abdul, C.-W. Chen, and C.-C. Wu, Biomedical Informatics, vol. 63, pp. 22–32, 2016.
“Recognition and evaluation of clinical section headings in [30] J. Dziadek, A. Henriksson, and M. Duneld, “Improving ter-
clinical documents using token-based formulation with minology mapping in clinical text with context-sensitive
conditional random fields,” BioMed Research International, spelling correction,” Studies in Health Technology and In-
vol. 2015, Article ID 873012, 10 pages, 2015. formatics, vol. 235, pp. 241–245, 2017.
[16] J. Liang, X. Xian, X. He et al., “A novel approach towards [31] Y. Luo, Ö. Uzuner, and P. Szolovits, “Bridging semantics and
medical entity recognition in Chinese clinical text,” Journal of syntax with graph algorithms-state-of-the-art of extracting
Healthcare Engineering, vol. 2017, Article ID 4898963, biomedical relations,” Briefings in Bioinformatics, vol. 18,
16 pages, 2017. no. 1, pp. 160–178, 2016.
[17] J. Xu, L. Gan, M. Cheng, and Q. Wu, “Unsupervised medical [32] J. R. Hobbs, “Resolving pronoun references,” in Reading in
entity recognition and linking in Chinese online medical text,” Natural Language Processing, B. Grosz, K. Sparck-Jones, and
Journal of Healthcare Engineering, vol. 2018, Article ID B. Webber, Eds., vol. 99, pp. 339–352, Morgan Kaufmann
2548537, 13 pages, 2018. Publishers, Los Altos, CA, USA, 1986.
[33] J. R. Tetrault, “A corpus-based evaluation of centering and Resources and Evaluation (LREC-2018), pp. 395–400, Miya-
pronoun resolution,” Computational Linguistics, vol. 27, no. 4, zaki, Japan, May 2018.
pp. 507–520, 2001. [49] S. Toldova, I. Azerkovich, A. Ladygina, A. Roitberg, and
[34] D. K. Byron, “Resolving pronominal references to abstract M. Vasilyeva, “Error analysis for anaphora resolution in
entities,” in Proceedings of the 40th Annual Meeting of the Russian: new challenging issues for anaphora resolution task
Association for Computational Linguistics (ACL), pp. 80–87, in a morphologically rich language,” in Proceedings of the
Philadelphia, PA, USA, July 2002. Workshop on Coreference Resolution Beyond OntoNotes,
[35] R. Mitkov and C. Hallett, “Comparing pronoun resolution CORBON@HLT-NAACL, pp. 74–83, San Diego, CA, USA,
algorithms,” Computational Intelligence, vol. 23, no. 2, June 2016.
pp. 262–297, 2007. [50] A. Khadzhiiskaia and A. Sysoev, “Coreference resolution for
[36] S. Lappin and H. J. Leass, “An algorithm for pronominal Russian: taking stock and moving forward,” in Proceedings of
anaphora resolution,” Computational Linguistics, vol. 20, 2017 Ivannikov ISPRAS Open Conference (ISPRAS), pp. 70–75,
no. 4, pp. 535–561, 1994. Moscow, Russia, November–December 2017.
[37] A. Zeldes and S. Zhang, “When annotation schemes change [51] L. Kučová and Z. Žabokrtský, “Anaphora in Czech: large data
rules help: a configurable approach to coreference resolution and experiments with automatic anaphora resolution,” in
beyond ontonotes,” in Proceedings of Workshop on Corefer- Text, Speech and Dialogue, pp. 93–98, Springer, Berlin,
ence Resolution Beyond OntoNotes (CORBON 2016), pp. 92– Germany, 2005.
[52] G. L. Nguy, V. Novák, and Z. Zabokrtský, “Comparison of
101, San Diego, CA, USA, June 2016.
classification and ranking approaches to pronominal
[38] N. Ge, J. Hale, and E. Charniak, “A statistical approach to
anaphora resolution in Czech,” in Proceedings of the 10th
anaphora resolution,” in Proceedings of the Sixth Workshop on
Annual Meeting of the Special Interest Group on Discourse and
Very Large Corpora, pp. 161–170, Montreal, Canada, August
Dialogue, SIGDIAL 2009, pp. 276–285, London, UK, Sep-
1998.
tember 2009.
[39] R. Mitkov, R. Evans, and C. Orasan, “A new, fully automatic
[53] M. Novák, “Coreference resolution system not only for
version of mitkov’s knowledge-poor pronoun resolution Czech,” in Proceedings of the 17th conference ITAT 2017:
method,” in Proceedings of the International Conference on Slovenskočeský NLP Workshop (SloNLP 2017), Praha, UK,
Intelligent Text Processing and Computational Linguistics, September 2017.
pp. 168–186, Springer, Budapest, Hungary, April 2002. [54] M. Novák, A. Nedoluzhko, and Z. Žabokrtský, “Projection-
[40] W. M. Soon, H. T. Ng, and D. C. Y. Lim, “A machine learning based coreference resolution using deep syntax,” in Pro-
approach to coreference resolution of noun phrases,” Com- ceedings of the 2nd Workshop on Coreference Resolution Be-
putational Linguistics, vol. 27, no. 4, pp. 521–544, 2001. yond OntoNotes (CORBON 2017), pp. 56–64, Valencia, Spain,
[41] J. R. Finkel and C. D. Manning, “Enforcing transitivity in April 2017.
coreference resolution,” in Proceedings of the 46th Annual [55] V. Žitkus, Lithuanian Coreference Corpus, CLARIN-LT
Meeting of the Association for Computational Linguistics on Digital Library, Lithuania, UK, 2018, http://hdl.handle.net/
Human Language Technologies: Short Papers, pp. 45–48, 20.500.11821/19.
Columbus, OH, USA, June 2008.
[42] S. Wiseman, A. M. Rush, and S. M. Shieber, “Learning global
features for coreference resolution,” in Proceedings of Con-
ference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies,
San Diego, CA, USA, June 2016.
[43] K. Lee, L. He, M. Lewis, and L. Zettlemoyer, “End-to-end
neural coreference resolution,” in Proceedings of 2017 Con-
ference on Empirical Methods in Natural Language Processing,
Association for Computational Linguistics, Brussels, Belgium,
October–November 2017.
[44] S. Žitnik, L. Šubelj, and M. Bajec, “SkipCor: skip-mention
coreference resolution using linear-chain conditional random
fields,” PLoS ONE, vol. 9, no. 6, Article ID e100101, 2014.
[45] A. Znotins and P. Paikens, “Coreference resolution for lat-
vian,” in Proceedings of the 9th International Conference on
Language Resources and Evaluation, LREC 2014, pp. 3209–
3213, Reykjavik, Iceland, May 2014.
[46] M. Ogrodniczuk and M. Kopeć, “Rule-based coreference
resolution module for polish,” in Proceedings of 8th Discourse
Anaphora and Anaphor Resolution Colloquium, DAARC 2011,
pp. 191–200, Faro, Portugal, October 2011.
[47] M. Kopec and M. Ogrodniczuk, “Creating a coreference
resolution system for polish,” in Proceedings of the Eighth
International Conference on Language Resources and Evalu-
ation (LREC), pp. 192–195, Istanbul, Turkey, May 2012.
[48] B. Nitoń, P. Morawiecki, and M. Ogrodniczuk, “Deep neural
networks for coreference resolution for Polish,” in Proceedings
of the Eleventh International Conference on Language
Hindawi
https://doi.org/10.1155/2019/4762490
Research Article
Brain Tissue Segmentation and Bias Field Correction of
MR Image Based on Spatially Coherent FCM with
Nonlocal Constraints
1,2
Jianhua Song and Zhe Zhang2
1
College of Physics and Information Engineering, Minnan Normal University, Zhangzhou 363000, China
2
Electronic Engineering College, Heilongjiang University, Harbin 150080, China
Correspondence should be addressed to Jianhua Song; 98dg_sjh@163.com
Received 27 September 2018; Accepted 11 February 2019; Published 3 March 2019
Copyright © 2019 Jianhua Song and Zhe Zhang. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Influenced by poor radio frequency field uniformity and gradient-driven eddy currents, intensity inhomogeneity (or bias field)
and noise appear in brain magnetic resonance (MR) image. However, some traditional fuzzy c-means clustering algorithms with
local spatial constraints often cannot obtain satisfactory segmentation performance. Therefore, an objective function based on
spatial coherence for brain MR image segmentation and intensity inhomogeneity correction simultaneously is constructed in this
paper. First, a novel similarity measure including local neighboring information is designed to improve the separability of MR data
in Gaussian kernel mapping space without image smoothing, and the similarity measure incorporates the spatial distance and
grayscale difference between cluster centroid and its neighborhood pixels. Second, the objective function with an adaptive
nonlocal spatial regularization term is drawn upon to compensate the drawback of the local spatial information. Meanwhile, bias
field information is also embedded into the similarity measure of clustering algorithm. From the comparison between the
proposed algorithm and the state-of-the-art methods, our model is more robust to noise in the brain magnetic resonance image,
and the bias field is also effectively estimated.
1. Introduction [2–4] and the fuzzy c-means (FCM) algorithm. Random field
is density-based unsupervised method where it finds the
Magnetic resonance image has been widely used in di- maximum likelihood estimate of the parameters from a
agnostic imaging for general check-up in clinical applica- given dataset. However random field algorithm has the
tion, especially the detection and diagnosis of brain diseases. disadvantages in high complexity and slow convergence and
The volume change for brain tissues often indicates various will drop into local optimization. FCM clustering is another
diseases [1], such as brain tumor, leukoencephalopathy, efficient method used in image segmentation because it has
olivopontocerebellar atrophy (OPCA), etc. Therefore, brain robust characteristics for ambiguity and can retain much
tissue segmentation of MR image has become a very im- more information than random field algorithm [5].
portant medical treatment step. However, brain MR image Therefore, FCM has been widely applied in different types of
has some lacks such as noise, intensity inhomogeneity, low image segmentation [6–8]. The neighboring pixels in an
contrast, the partial volume effect, and so on, which brings image are highly correlated, i.e., the pixels in the immediate
serious obstacles to segment the brain MR images. To this neighborhood possess nearly the same feature data.
end, the multitudinous brain MR image segmentation Therefore, the spatial relationship of neighboring pixels is an
methods have been put forward by using the theory of fuzzy important characteristic that can be of great aid in imaging
set, random field, and level set. segmentation. However, the conventional FCM algorithm
Currently, there are two popular methods-based models does not fully utilize this spatial information. Pedrycz and
for medical image segmentation: the random field theory Waletzky [9] took advantage of the available classified
information and actively applied it as part of their optimi- spatially coherent FCM with nonlocal constraints is proposed.
zation procedures. Szilagyi et al. [10] proposed the enhanced In this model, firstly, both the local constraint term and
FCM (EnFCM) algorithm to accelerate the image segmen- nonlocal regularization term about brain MR image are in-
tation process in which the pixels of an image are replaced the corporated into the objective function, and an adaptive control
gray-level histogram and the statistical number and calcula- factor is used to maintain the balance between them. Secondly,
tion are much smaller than FCM. In order to further reduce the similarity measure is designed in Gaussian kernel mapping
the computation time and improve the parameter inflexibility, space without image filtering, and the detail information and
Cai et al. [11] presented a fast generalized FCM (FGFCM) the edge of the image can be preserved well. Meanwhile, bias
method, and FGFCM introduced a flexible locality factor Sij field model is also embedded into the objective function of
incorporating simultaneously both the gray-level difference clustering algorithm. Therefore, after the intensity in-
and spatial distance in a local window. Ji et al. [12] proposed a homogeneity of the MR image is corrected, the segmentation
robust spatially constrained fuzzy c-means (RSCFCM) algo- accuracy is improved significantly. Experiments demonstrate
rithm for brain MR image segmentation. First, a spatial factor that this algorithm can not only effectively estimate the bias
is constructed based on the posterior probabilities and prior field of MR image but also has stronger antinoise ability.
probabilities and takes the spatial direction into account.
Second, the negative log-posterior is utilized as dissimilarity 2. Preliminary Theory
function by taking the prior probabilities into account.
FCM with spatial constraint and its variants greatly im- 2.1. Fuzzy Clustering Algorithms. Let X � 􏼈x1 , x2 , . . . , xN 􏼉
proved the antinoise performance compared with FCM, but denote an image with N pixels, where xk is gray value of the
when the noise is very serious in the image, the performance kth pixel in the image. FCM clustering aims at partitioning X
of the algorithm may be worse. Therefore, the nonlocal spatial into c clusters by minimizing the following objective function:
c N �� 2
information was often used and incorporated into the dis- JFCM � 􏽘 􏽘 um � �
ik �xk − vi � , (1)
tance metric of FCM in recent years [13–16]. Zhao [14] i�1 k�1
brought in a nonlocal adaptive regularization term in its
energy function, and the control factor is adaptive determined where vi denotes the ith cluster prototype, uik denotes the
to adjust the balance of the objective function. Feng et al. [15] membership degree of xk belonging to ith cluster and fol-
proposed a FCM method with specific nonlocal information lows 􏽐ci�1 uik � 1, c denotes the number of centers, ‖ · ‖
for the segmentation of synthetic aperture radar (SAR) image. denotes the Euclidean norm, and the parameter m is a
Ma et al. [16] proposed a modified FGFCM approach by weight exponent on each fuzzy membership that determines
introducing nonlocal constraint term, and local distance the amount of fuzziness of the resulting partition.
metric and nonlocal distance metric are used, respectively, in Ahmed et al. proposed a modification to FCM objective
its objective function. By introducing nonlocal constraint function by introducing a term that allows the labeling of a
term, the features of image can be used more comprehensively pixel to be influenced by the labels in its immediate
and effectively, and the robustness to noise of FCM-based neighborhood [23]. This effect acts as a regularizer and biases
algorithm is significantly improved. However, there generally the solution toward piecewise homogeneous labeling. It
exists intensity inhomogeneity in brain MR images. There- proved useful in segmenting images corrupted by noise. The
fore, it is necessary to further design relevant algorithms to modified objective function is given by
correct the intensity inhomogeneity. Sled et al. designed a set c N �� 2 α c N m ��
of software package for the estimation of bias field [17], and JBCFCM � 􏽘 􏽘 um � � ��xr − vi ��2 ,
ik �xk − vi � + 􏽘􏽘u 􏽘
the characteristic of the method is nonparametric non- i�1 k�1
NR i�1 k�1 ik r∈N
k
uniform intensity normalization or N3 for short; the distri- (2)

bution of the true tissue intensities can be achieved by an
iterative method. Tustison et al. [18] improved the N3 al- where xr represents the neighbor voxels of xk and NR and
gorithm based on modified B-spline approximation and hi- Nk stand for the number of voxels in the neighborhood of
erarchical optimization algorithm (called N4ITK); N4ITK can the kth voxel. The parameter α controls the intensity of the
also automatically perform without the priori knowledge. neighboring effect. One disadvantage of BCFCM is that the
Liew and Hong Yan [19] introduced a spatial constraint to a neighborhood labeling is computed in each iteration step,
fuzzy cluster method where the inhomogeneity field was something that is very time consuming.
modeled by a B-spline surface. The spatial pixel connectivity
was implemented by a dissimilarity index, which enforced the 2.2. Spatially Coherent Fuzzy c-Means Clustering (SCFCM).
connectivity constraint only in the homogeneous areas. Ji In view of some drawbacks of standard FCM algorithm,
et al. proposed the modified possibilistic FCM (MPFCM) a modified scheme is proposed by Despotović et al. [24].
algorithm for bias field [20], generalized rough fuzzy c-means The similarity measure Dik � ‖xk − vi ‖2 is replaced by D∗ik �
(GRFCM) algorithm, [21] and fuzzy local Gaussian mixture (1 − Sik )‖xk − vi ‖2 introducing a weight factor Sik , and the
model (FLGMM) for brain MR image segmentation [22], objective function is
respectively. Those methods can estimate bias field and
c N �� 2
segment the MR images simultaneously.
JSCFCM � 􏽘 􏽘 um � �
ik 1 − Sik 􏼁�xk − vi � , (3)
In this paper, a brain tissue classification and intensity i�1 k�1
inhomogeneity correction model of MR image based on
where Sik is a weight factor including both local spatial in- c N �� 2
formation and grayscale difference and is designed as follows: JCLIC � 􏽘 􏽘 um � �
ik 􏽘 K(r − k)�xk − br vi � , (7)
i�1 k�1 r∈Ωk
􏽐r∈Ωk uir akr d−1
kr
Sik � , (4)
􏽐r∈Ωk akr d−1 where br vi is the clustering prototype with bias field in the
kr
selective region Ωk , K(r − k) is the weight of a truncated
where Ωk denotes a local neighboring window around xk , uir Gaussian kernel allocated for the intensity xk , and the weight
denotes the membership degree of neighborhood pixels parameter can be defined as
belonging to the ith cluster, akr � |xk − xr | is the absolute ⎪
⎧ 1 −|y|2 /2σ 2
⎪ for |y| ≤ ρ,
intensity difference between the study pixel and its neighbor,
􏽱�� ⎨ ae
⎪ ,
K(y) � ⎪ (8)
dkr � (pk − pr )2 + (qk − qr )2 is the Manhattan distance ⎪
⎪
⎩
between the coordinates of pixel xk and xr , and (pk , qk ) and 0, else,
(pr , qr ) denote the coordinates xk and xr in the image,
respectively. By minimizing equation (3) by Lagrangian where σ denotes the standard deviation, a denotes a nor-
multiplier approach, uik and vi can be derived as shown in malization factor to standardize Gaussian kernel, and ρ
the following equation: denotes a radius to measure the size of the local region.
�� 2 −1/(m−1) In the CLIC model, intensity inhomogeneity of the MR
􏼒 1 − Sik 􏼁��xk − vi �� 􏼓 image can be effectively corrected and can reduce the
uik � misclassification rate, but there are some drawbacks in
�� 2 −1/(m−1) ,
􏽐cl�1 􏼒 1 − Slk 􏼁��xk − vl �� 􏼓 CLIC. When computing the distance metric between the
(5) central pixel and its surrounding pixels in a local region, the
model only used the local neighborhood information of the
􏽐N m
k�1 uik xk pixel without considering the global structure information of
vi � N .
􏽐k�1 umik the entire image. As a result, the antinoise performance of
this algorithm is unsatisfactory.
Compared with the FCM, this algorithm has two ad-
vantages: firstly, it enhances the robustness to all kinds of
noise. The constraint item of neighborhood information is 3. Proposed Method
included into the similarity measure, so as to effectively The standard FCM algorithm has the shortcoming of being
utilize the local information of the image. Secondly, it sensitive to noise. Though, the modified FCM algorithms are
considers anisotropic neighborhood information, and more
improved by adding the spatial information, it is difficult to
details and edges information can be preserved. However,
get a satisfied segmentation result for noise robustness.
the influence of bias field for MR images in segmentation
Therefore, an improved FCM approach based on CLIC and
algorithm is not mentioned.
SCFCM is proposed; its objective function is constructed
according to the local constraint term and global regulari-
2.3. Coherent Local Intensity Clustering Model. Bias field of zation term; the similarity measure including local neigh-
the MR image usually embodies slowly and smoothly boring information is designed in Gaussian kernel mapping
varying for the pixel grayscale of the local region across an space, and brain tissue classification and intensity in-
image. Meanwhile, in a neighboring local window of the homogeneity correction can be realized simultaneously.
image, bias field can be approximately considered as a
constant. Therefore, the most popular model can be de-
scribed in equation (6) [25]; let Y denote the observed image, 3.1. Nonlocal Weighted Constraint. In a discrete noisy image
b denote the unknown bias field, X denote the true image to X � 􏼈x1 , x2 , . . . , xN 􏼉, xk is the kth pixel and yk is its nonlocal
be restored, and n denote the additive noise. weighted average value. The derivation method of the
nonlocal weighted average can be acquired in [27], and the
Y � bX + n. (6)
mathematical expression of yk is
In the observed image, noise n is often assumed to obey yk � 􏽘 wkl xl ,
Gaussian distribution with zero mean and variance σ 2n , and (9)
l∈wuk
the gray level value of the true image approximately takes a
constant in a local window. Therefore, the gray level of the where wuk indicates a search region of radius u around xk , wkl
observed image can be approximated to obey Gaussian denotes nonlocal weight coefficient depending on similarity
distribution with mean bX and variance σ 2n . In coherent local measure between xk and its neighboring pixels xl in window
intensity clustering (CLIC) model [26], a novel metric in- wuk , and wkl satisfies the constraint conditions 0 ≤ wkl ≤ 1 and
troducing spatially coherent local intensity convergence 􏽐l wkl � 1. The weight coefficient wkl is computed as follows:
criterion for bias field estimation and image segmentation 1 2
simultaneously is proposed. A Gaussian kernel weight pa- wkl � e −‖x(Nk )−x(Nl )‖2,σ /h􏼁 , (10)
Zk
rameter K(r − k) is introduced into the similarity measure of
each pixel gray level xk and its neighbor pixels xr , and the where x(Nk ) and x(Nl ) denote the grayscale vectors in the
objective function of CLIC is square neighborhood Nk and Nl of radius s around xk and
xl , respectively, and ‖x(Nk ) − x(Nl )‖22,σ is the weighted where uik is the membership degree of xk belonging to the
Euclidean distance between x(Nk ) and x(Nl ); its expression ith cluster, Ωk denotes a local square region of the radius s
is defined in equation (11). σ is the same as in equation (8), around the center xk , bk vi is the ith cluster center in Ωk , yk
and h denotes a control factor to adjust the variation of the denotes the nonlocal weighted average value of xk , Sik de-
similarity measure wkl . notes the weight factor of local neighborhood information in
(2s+1) 2 equation (4), and the definitions of K(r − k) and bk are the
��
��x Nk 􏼁 − x Nl 􏼁��2 � 􏽘 σ p xp Nk 􏼁 − xp Nl 􏼁􏼁2 , (11) same as equation (7). αk is a trade-off weight factor to adjust
2,σ
p�1 the balance of local neighborhood information and nonlocal
constraints information, and the definition of parameter
where xp (Nk ) is the pth pixel in the grayscale vectors x(Nk ) αk is
and σ p is defined as follows:
s 1
p 1 αk � , (14)
σ � 􏽘 2 , (12) 1 + var(x)/(x)2
v�max(d,1) (2v + 1) s
where x denotes the gray level mean of all pixels in the local
where yp � mod(p, (2s + 1)) and zp � floor(p, (2s + 1))+ region Ωk , and var(x) denotes the variance of pixel gray
1, (yp , zp ) denote the coordinates of the pth component in a values in the same window. The larger the αk value is, the
preselected region and d � max(|yp − s − 1|, |zp − s − 1|). smaller the influence of the noise is. The factor αk can be
obtained adaptively with the change of the local window Ωk
3.2. Objective Function. In order to correct bias field and without being given in advance.
classify the brain tissues simultaneously, the modified ob-
jective function-incorporated local constraint term and
nonlocal regularization term is as follows: Theorem. Assume 􏽐ci�1 uik � 1, 0 ≤ uik ≤ 1, and m > 1. On the
c N
basis of Lagrange multiplier approach, equation (13) is
�� 2
Jm � 􏽘 􏽘 um � � minimized with respect to uik , vi and bk can be derived as
ik 􏽘 K(r − k)􏼢αk 1 − Sik 􏼁�xk − bk vi �
i�1 k�1 r∈Ωk shown in the following equation:
(13)
�� 2
+ 1 − αk 􏼁��yk − bk vi �� 􏼣,
�� 2 �� 2 −1/(m−1)

􏽐r∈Ωk K􏼔αk 1 − Sik 􏼁��xk − bk vi �� + 1 − αk 􏼁��yk − bk vi �� 􏼕
uik � �� 2 �� 2 −1/(m−1) ,
􏽐cl�1 􏼒􏽐r∈Ωk K􏼔αk �
1 − Slk 􏼁�xk − bk vl � + 1 − αk 􏼁�yk − bk vl �� 􏼕􏼓
� �
􏽐N m
k�1 􏽐r∈Ωk Kbk 􏼂αk 1 − Sik 􏼁xk + 1 − αk 􏼁yk 􏼃uik
(15)
vi � ,
􏽐N 2 m
k�1 􏽐r∈Ωk Kbk 1 − αk Sik 􏼁uik
􏽐ci�1 􏽐r∈Ωk Kvi 􏼂αk 1 − Sik 􏼁xk + 1 − αk 􏼁yk 􏼃um

ik
bk � .
􏽐ci�1 􏽐r∈Ωk Kv2i 1 − αk Sik 􏼁um
ik
Proof. According the method of Lagrange multiplier, equa- where λk is the Lagrange multiplier of the constraint con-
tion (13) can be converted to unconstrained optimization dition 􏽐ci�1 uik � 1, by computing the partial derivatives of
problem: polynomial L in regard to uik and λk , respectively, and set
c N zL/zuik � 0, zL/zλk � 0, as shown in the following equation:
L uik , vi , bk , λk , ck 􏼁 � 􏽘 􏽘 um
ik 􏽘 αk K(r − k) zL �� 2
i�1 k�1 r∈Ωk � mum−1ik 􏽘 K􏼢αk 1 − Sik 􏼁��xk − bk vi ��
zuik
�� 2 r∈Ωk
· 􏼢 1 − Sik 􏼁��xk − bk vi �� (17)
(16) �� 2
�� 2 + 1 − αk 􏼁��yk − bk vi �� 􏼣 − λk � 0,
+ 1 − αk 􏼁��yk − bk vi �� 􏼣
N c c
⎣1 − 􏽘 uik ⎤⎦, zL
+ 􏽘 λk ⎡ � 1 − 􏽘 uik � 0. (18)
k�1 i�1 zλk i�1
The following equation can be obtained by mathematical λ

1/m−1 c � �
derivation of equation (17): 􏼠 k􏼡 � 􏽘⎛⎝ 􏽘 K􏼢α 1 − S 􏼁��x − b v ��2
k ik k k i
m i�1 r∈Ωk
1/m−1
⎜
⎛ λk ⎟
⎞ (20)
uik � ⎜
⎜
⎜
⎝ �� 2 ��
⎟
��2 ⎟
⎟
⎠ . −(1/m−1)
� � � �
m􏽐r∈Ωk K􏼔αk 1 − Sik 􏼁�xk − bk vi � + 1 − αk 􏼁�yk − bk vi � 􏼕 �� 2
+ 1 − αk 􏼁��yk − bk vi �� 􏼣⎞⎠ .
(19)
Substituting equation (19) into equation (17), we obtain And then substituting equation (20) into equation (19),
the following equation: the following equation can be obtained:
�� 2 �� 2 −(1/m−1)

􏽐r∈Ωk K􏼔αk 1 − Sik 􏼁��xk − bk vi �� + 1 − αk 􏼁��yk − bk vi �� 􏼕
uik � �� 2 �� 2 −(1/m−1) . (21)
􏽐cl�1 􏼒􏽐r∈Ωk K􏼔αk �
1 − Slk 􏼁�xk − bk vl � + 1 − αk 􏼁�yk − bk vl �� 􏼕􏼓
� �
Similarly, set zL/zvi � 0, that is estimation and antinoise performance analysis for the brain
N MR images are the main experimental contents. For the
zL
� 􏽘 um ⎛
⎝ 􏽘 Kb 􏼔α 1 − S 􏼁 x − b v 􏼁
k k ik k k i
experiments in the following sections, the related parameter
zvi k�1 ik r∈Ω values are fuzzy exponential m � 2, the stop criterion
k
(22) ε � 0.001, 3 × 3 neighborhood window, and the radius u �
10 of the search window.
⎠ � 0.
+ 1 − αk 􏼁 yk − bk vi 􏼁􏼕⎞
4.1. Bias Field Correction

The following equation can be obtained from equation
(22) by mathematical derivation: 4.1.1. MR Image Database. Intensity inhomogeneity is one
􏽐N m of the problems in interfering brain MR image segmenta-
k�1 􏽐r∈Ωk Kbk 􏼂αk 1 − Sik 􏼁xk + 1 − αk 􏼁yk 􏼃uik
vi � . (23) tion; in the experiments of bias field correction, the dataset is
􏽐N 2 m
k�1 􏽐r∈Ωk Kbk 1 − αk Sik 􏼁uik from a simulated brain database (SBD) : BrainWeb [28] in
which the brain MR images have three types: T1-, T2-, and
We adopt the same mathematical derivation process to
proton density- (PD-) weighted 3D data volumes. In Fig-
estimate bias field bk , for fixed uik and λk , and computing the
ure 1, the T1-weighted normal brain MR images with 181 ×
partial derivative of L with respect to bk , set zL/zbk � 0, that
217 × 181 cubic voxels, 1 mm slice thickness, 40% intensity
is
nonuniformity, and 3% noise are used to test; all the skulls
zL c and blood vessels are already stripped ahead of image
� 􏽘 um ⎛ 􏽘 K􏼔αk 1 − Sik 􏼁􏼐vi xk − v2 bk 􏼑
⎝ processing, and the image is segmented into four regions:
zbk i�1 ik r∈Ω i
k white matter (WM), gray matter (GM), cerebrospinal fluid
(24) (CSF), and background.
⎠ � 0,
+ 1 − αk 􏼁􏼐vi yk − v2i bk 􏼑􏼕⎞
4.1.2. Experimental Results. Figure 1 shows the results of bias
bk can be obtained from equation (24). field correction and segmentation for three brain MR images.
Figure 1(a) shows the brain slice images from three different
􏽐ci�1 􏽐r∈Ωk Kvi 􏼂αk 1 − Sik 􏼁xk + 1 − αk 􏼁yk 􏼃um
ik directions: transaxial mode, sagittal mode, and coronal mode.
bk � . (25)
􏽐ci�1 􏽐r∈Ωk Kv2i 1 − αk Sik 􏼁um
ik
Figure 1(b) shows the estimated bias field, Figure 1(c) shows
the final segmentation results, and Figure 1(d) shows the
The theorem proves to be completed. corrected images after removed bias field. Figure 2 shows the
Finally, the framework of the proposed algorithm can be histogram comparison of original MR image and bias cor-
summarized in Table 1. □ rected images corresponding to three images in Figure 2.
From Figures 1 and 2, three brain tissues are more homo-
geneous after bias field correction; each brain tissue ap-
4. Experimental Results and Analysis proximately obeys Gaussian distribution with different mean
and variance, and WM, GM, and CSF can be clearly dis-
In this section, several classical algorithms for intensity tinguished. In addition, the histogram distribution of cor-
inhomogeneity correction and brain image segmentation rected MR image is more reasonable, from which we can see
are selected as the reference for comparison; bias field three approximative peaks representing three brain tissues.
Table 1: The basic flow of image segmentation and bias field estimation.
Step 1. Set the number of cluster c, the exponent of fuzziness m, stopping condition of the iteration ε > 0, the radius s of local neighborhood
window, and the radius u of search window in equation (11).
Step 2. Set the iteration initial value t � 0 and initialize bias field b0 � 1 and the center of cluster V1 � 􏼈v11 , v12 , . . . , v1c 􏼉.
Step 3. Calculate nonlocal weight coefficient wtkl in equation (8) and then obtain the nonlocal weighted value ytk by equation (7).
Step 4. Update the membership degree utik by equation (14).
Step 5. Update the center of clustering vti by equation (14).
Step 6. Calculate the estimated bias field btk by equation (15).
Step 7. If satisfying max‖Vt+1 − Vt ‖ < ε, then terminate iteration; otherwise, go to step 3 and set t � t + 1.
(a) (b) (c) (d)
Figure 1: Intensity inhomogeneity correction of brain MR images: (a) original MR images, (b) the estimated bias field, (c) segmentation
results, and (d) the corrected MR images.
To further validate the performance of bias field cor- algorithms of bias field estimation are more or less effective.
rection, three bias correction algorithms including BCFCM However, our method is more reasonable than other three
[23], CLIC [26], and N4ITK [18] are chosen as comparative algorithms from Figure 4. Because the histogram normally
methods, as shown in Figure 3. Figure 3(a) is a T1-weighted should have three peaks corresponding to WM, GM, and
transaxial slice of normal brain MR image with 40% spatial CSF; the peak value of CSF is minimal according to tissue
inhomogeneity; its slice thickness is 1 mm and the noise level volume, and the gray level’s mean and variance of each tissue
is about 2%. Figures 3(b) and 3(c) are the obtained bias are also obviously different.
inhomogeneity and the amendatory MR images by BCFCM,
CLIC, N4ITK, and our method, respectively. Figure (4)
shows the histograms of the image with spatial in- 4.2. Antinoise Performance. In the third experiment, first of
homogeneity and the corrected images in Figure 3. It is can all, the 104th transaxial slice of simulated T1-weighted brain
be seen that the distribution of pixel gray level of blue line is MR image with 1 mm slice thickness and 7% Gaussian white
more accurate than red dotted line; it indicates that all the noise is used to analyze the robustness to the noise, and we
500 500 500
400 400 400
300 300 300
200 200 200
100 100 100
0 0 0
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
Original image Original image Original image
Corrected image Corrected image Corrected image
(a) (b) (c)
Figure 2: The histogram comparison of original images and corrected images: (a) transaxial slice image, (b) sagittal slice image, and (c)
coronal slice image.
(a) (b) (c)
Figure 3: The comparison of bias field estimation by four algorithms: (a) original images, (b) the estimated bias field, and (c) the corrected
images.
select four algorithms: standard FCM, BCFCM [23], CLIC and the corresponding classification results by FCM,
[26], and SCFCM [24] as the compared algorithms. The BCFCM, CLIC, SCFCM, and the proposed method are
segmentation results are exhibited in Figure 5; Figure 5(a) is shown in the Figures 5(b)–5(f ), respectively. The segmen-
a 2D transaxial slice image corrupted by 7% Gaussian noise, tation results of FCM, BCFCM, and CLIC are very poor
500 500 500 500
400 400 400 400
300 300 300 300
200 200 200 200
100 100 100 100
0 0 0 0
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
Original image Original image Original image Original image

BCFCM CLIC N4ITK The proposed method
(a) (b) (c) (d)
Figure 4: The histogram comparison of original images and corrected images: (a) BCFCM, (b) CLIC, (c) N4ITK, and (d) our method.
(a) (b) (c)
(d) (e) (f) (g)
Figure 5: Segmentation results of the 104th transaxial slice by the five algorithms: (a) the noisy MR images, (b) FCM, (c) BCFCM, (d) CLIC,
(e) SCFCM, (f ) our method, and (g) ground truth.
because many pixels are misclassified; the segmentation In order to further evaluate and compare the antinoise
result by SCFCM is better than FCM, BCFCM, and CLIC; ability of aforementioned five fuzzy clustering algorithms,
however, there are still some noisy points that need to be we choose a brain slice image with 14% additive Gaussian
removed. It can be observed that the presented method has white noise as the test object, as shown in Figure 7.
more superior segmentation effect than four classical al- Figure 7(a) is the noisy MR image with bias field,
gorithms and can effectively eliminate the influence of the Figures 7(b)–7(f ) are the binary images of CSF, WM, and
noise. Furthermore, the ability of detail and edge preser- GM after the image is segmented by five algorithms, re-
vation is also compared and analyzed for five algorithms; we spectively, and Figure 7(g) is the ground truth. It is clearly
select a local region of the MR image to observe by enlarging seen that the extraction result of each brain tissue of the
3 times, and the detailed images are presented in Figure 6; it proposed algorithm significantly outperformed the other
is clearly seen that Figure 6(f ) is most similar subimage with algorithms and effectually overcame the disadvantageous
the ground truth in Figure 6(g), and the vast majority of defects of intensity inhomogeneity and noise. At the same
image details and edge are completely preserved. time, an objective evaluation index JS (Jaccard similarity)
(a) (b) (c)
(d) (e) (f ) (g)
Figure 6: The detailed comparison of the enlarged local region of brain MR image: (a) the noisy MR images, (b) FCM, (c) BCFCM, (d) CLIC,
(e) SCFCM, (f ) our method, and (g) ground truth.
(a) (b)
(c) (d)
(e) (f )
(g)
Figure 7: Brain tissue classification results of the five methods on the noisy MR image: (a) original MR image, (b) FCM, (c) BCFCM, (d)
CLIC, (e) SCFCM, (f ) our method, and (g) ground truth.
1.0 1.0 1.0

0.9 0.9 0.9
0.8 0.8 0.8

0.7 0.7
JS (WM)
JS (CSF)
JS (GM)
0.7
0.6 0.6
0.6
0.5
0.5
0.5 0.4
0.4
0.4 0.3
0.3
0.3 0.2
5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30
Noise level (%) Noise level (%) Noise level (%)
FCM SCFCM FCM SCFCM FCM SCFCM

BCFCM Our method BCFCM Our method BCFCM Our method
CLIC CLIC CLIC
(a) (b) (c)
Figure 8: JS comparisons of the brain tissue segmentation with different noise level by five algorithms: (a) JS values of WM, (b) JS values of
GM, and (c) JS values of CSF.
1.00 0.99
0.99
0.98
0.98
0.97
JS
JS
0.97
0.96
0.96
0.95 0.95
4 6 8 10 12 4 6 8 10 12
u u
s=1 s=5 s=1 s=5
s=3 s=7 s=3 s=7
(a) (b)
Figure 9: JS value under different search window radius u and square neighborhood radius s: (a) curves on the image with 1% noise level and
(b) curves on the image with 3% noise level.
[29] is adopted for comparison and quantitative analysis on Figure 8(a)–8(c). It is clearly shown that the presented
the different level noise, JS is given as approach has better matching degree with the ground truth
􏼌􏼌 􏼌 and higher accuracy rate than other four clustering methods.
􏼌􏼌S1 ∩ S2 􏼌􏼌􏼌 Then, we evaluate the effect of the search window radius
JS S1 , S2 􏼁 � 􏼌􏼌􏼌 􏼌􏼌, (26)
􏼌S1 ∪ S2 􏼌􏼌 u and the square neighborhood radius s on the performance
of the proposed method. Here, we test u and s on the sets {4,
where S1 represents a set of pixels of the segmented region by 6, 8, 10, 12} and {1, 3, 5, 7}, respectively. In this experiment,
a clustering algorithm, S2 denotes the set of pixels of the the tested images perform 8 independent runs of the al-
corresponding region acquired from the ground truth, ∩ gorithm under each pair (u, s), and the noise level is 1% and
denotes the intersection operation, and ∪ denotes the union 3%, respectively. Under each s value, the average JS curve of
operation. As a quantitative evaluation index, the values of JS the algorithm with the increase of u value is shown in
belong to the interval of [0, 1], and the higher the JS value is, Figure 9. It can be found from Figure 9 that the algorithm
the better the segmentation performance is. We selected 15 under s � 3 and u � 10 can obtain satisfying performance on
noisy brain MR images with 20% intensity nonuniformity as the noisy images.
the experimental objects and the noise level from 5% to 30%. In the aforementioned sections, the model is applied in
These images are segmented three regions: WM, GM, and the synthetic brain MR images. Next, this model is also
CSF by FCM, BCFCM, CLIC, SCFCM, and our method, applied to the real clinical images with noise. We selected
respectively; JS values comparison results are shown in three normal MR slice images from transaxial, coronal, and
(a)
(b)
(c)
(d)
(e)
Figure 10: Experimental results on real MR image: (a) three original images, (b) BCFCM, (c) HMRF-EM, (d) SCFCM, and (e) our method.
sagittal views, and these MR images are obtained from the T1-weighted brain MR slice images; the left image is a
Whole Brain Atlas clinical MR image database by the transaxial slice, the right image is a coronal slice, and the
Harvard medical school [30]. Figure 10(a) shows three 2D middle image is a sagittal slice. The segmentation results of
brain slice images are given in Figure 10(b)–10(e) by image segmentation,” Medical Image Analysis, vol. 16, no. 8,
BCFCM, HMRF-EM, SCFCM, and our method, re- pp. 1477–1489, 2012.
spectively. From the experimental results, it is obvious that [4] M. Xie, J. Gao, C. Zhu, and Y. Zhou, “A modified method for
the proposed method can effectively segment each brain MRF segmentation and bias correction of MR image with
tissues as well as preserve more detail information of the intensity inhomogeneity,” Medical & Biological Engineering &
Computing, vol. 53, no. 1, pp. 23–35, 2015.
original MR image. Furthermore, the experimental results of
[5] M. A. Balafar, “Fuzzy C-mean based brain MRI segmentation
brain tissues in real MR images also further prove the ro- algorithms,” Artificial Intelligence Review, vol. 41, no. 3,
bustness to noise of the proposed method. pp. 441–449, 2014.
[6] Z. Ji, Y. Xia, Q. Chen, Q. Sun, D. Xia, and D. D. Feng, “Fuzzy
5. Conclusion c-means clustering with weighted image patch for image
segmentation,” Applied Soft Computing, vol. 12, no. 6,
Brain MR imaging has wide clinical application as an ef- pp. 1659–1667, 2012.
fective medical imaging diagnostic technique; however, the [7] J. Aparajeeta, P. K. Nanda, and N. Das, “Modified possibilistic
real brain MR images often suffer from some interference fuzzy C-means algorithms for segmentation of magnetic
such as noise, intensity inhomogeneity, and low contrast. resonance image,” Applied Soft Computing, vol. 41, pp. 104–
Therefore, a brain tissue classification and nonuniformity 119, 2016.
field correction scheme in MR images based on spatially [8] J. Zheng, D. Zhang, K. Huang, and Y. Sun, “Adaptive image
segmentation method based on the fuzzy c-means with spatial
coherent FCM with nonlocal constraints is proposed in our
information,” IET Image Processing, vol. 12, no. 5, pp. 785–
study. The available information including local adjacent 792, 2018.
constraint and nonlocal global information of brain MR [9] W. Pedrycz and J. Waletzky, “Fuzzy clustering with partial
image is fully used in our model, and the similarity measure supervision,” IEEE Transactions on Systems, Man and Cy-
is designed in Gaussian kernel mapping space. Furthermore, bernetics, Part B (Cybernetics), vol. 27, no. 5, pp. 787–795,
the algorithm corrects the bias field of the MR image and 1997.
improves its antinoise performance. Several experiments on [10] L. Szilagyi, Z. Benyo, S. M. Szilagyi et al., “MR brain image
the simulated brain MR images and real brain MR images segmentation using an enhanced fuzzy C-means algorithm,”
have demonstrated that the proposed model can effectually in Proceedings of International conference of the IEEE engi-
overcome the effects of the noise while estimating the bias neering in medicine and biology society, vol. 1, pp. 724–726,
field existing in brain MR images. Cancun, MX, USA, September 2003.
[11] W. Cai, S. Chen, and D. Zhang, “Fast and robust fuzzy
c-means clustering algorithms incorporating local in-
Data Availability formation for image segmentation,” Pattern Recognition,
vol. 40, no. 3, pp. 825–838, 2007.
The data used to support the findings of this study are [12] Z. Ji, J. Liu, G. Cao, Q. Sun, and Q. Chen, “Robust spatially
available from the corresponding author upon request. constrained fuzzy c-means algorithm for brain MR image
segmentation,” Pattern Recognition, vol. 47, no. 7, pp. 2454–
Conflicts of Interest 2466, 2014.
[13] Y. Chen, J. Li, H. Zhang et al., “Non-local-based spatially
The authors declare that there are no conflicts of interest constrained hierarchical fuzzy C-means method for brain
regarding the publication of this paper. magnetic resonance imaging segmentation,” IET Image
Processing, vol. 10, no. 11, pp. 865–876, 2017.
[14] F. Zhao, “Fuzzy clustering algorithms with self-tuning non-
Acknowledgments local spatial information for image segmentation,” Neuro-
computing, vol. 106, no. 6, pp. 115–125, 2013.
This work was partially supported by the Education and [15] J. Feng, L. C. Jiao, X. Zhang, M. Gong, and T. Sun, “Robust
Teaching Reform Project of Undergraduate Colleges and non-local fuzzy c-means algorithm with edge preservation for
Universities in Fujian Province (Grant no. FBJG20180015) SAR image segmentation,” Signal Processing, vol. 93, no. 2,
and Principal Fund Project of Minnan Normal University pp. 487–499, 2013.
(Grant no. KJ18010). [16] J. Ma, D. Tian, M. Gong, and L. Jiao, “Fuzzy clustering with
non-local information for image segmentation,” International
References Journal of Machine Learning and Cybernetics, vol. 5, no. 6,
pp. 845–859, 2014.
[1] S. V. Srinivasan, K. Narasimhan, R. Balasubramaniyam, and [17] J. G. Sled, A. P. Zijdenbos, and A. C. Evans, “A nonparametric
S. Rishi Bharadwaj, “Diagnosis and segmentation of brain method for automatic correction of intensity nonuniformity
tumor from MR image,” in Advances in Intelligent Systems in MRI data,” IEEE Transactions on Medical Imaging, vol. 17,
and Computing, pp. 687–693, Springer, Berlin, Germany, no. 1, pp. 87–97, 1998.
2015. [18] N. J. Tustison, B. B. Avants, P. A. Cook et al., “N4ITK: im-
[2] M. A. Balafar, A. R. Ramli, and S. Mashohor, “Brain magnetic proved N3 bias correction,” IEEE Transactions on Medical
resonance image segmentation using novel improvement for Imaging, vol. 29, no. 6, pp. 1310–1320, 2010.
expectation maximizing,” Neurosciences, vol. 16, no. 3, [19] A. W.-C. Liew and H. Hong Yan, “An adaptive spatial fuzzy
pp. 242–247, 2011. clustering algorithm for 3-D MR image segmentation,” IEEE
[3] J. P. Monaco and A. Madabhushi, “Class-specific weighting Transactions on Medical Imaging, vol. 22, no. 9, pp. 1063–
for markov random field estimation: application to medical 1075, 2003.
[20] Z.-X. Ji, Q.-S. Sun, and D.-S. Xia, “A modified possibilistic
fuzzy c-means clustering algorithm for bias field estimation
and segmentation of brain MR image,” Computerized Medical
Imaging and Graphics, vol. 35, no. 5, pp. 383–397, 2011.
[21] Z. Ji, Q. Sun, Y. Xia, Q. Chen, D. Xia, and D. Feng, “Gen-
eralized rough fuzzy c-means algorithm for brain MR image
segmentation,” Computer Methods and Programs in Bio-
medicine, vol. 108, no. 2, pp. 644–655, 2012.
[22] Z. Ji, Y. Xia, Q. Sun et al., “Fuzzy local Gaussian mixture
model for brain MR image segmentation,” IEEE Transactions
on Information Technology in Biomedicine, vol. 16, no. 3,
pp. 339–347, 2012.
[23] M. N. Ahmed, S. M. Yamany, N. Mohamed, A. A. Farag, and
T. Moriarty, “A modified fuzzy C-means algorithm for bias
field estimation and segmentation of MRI data,” IEEE
Transactions on Medical Imaging, vol. 21, no. 3, pp. 193–199,
2002.
[24] I. Despotovic, E. Vansteenkiste, and W. Philips, “Spatially
coherent fuzzy clustering for accurate and noise-robust image
segmentation,” IEEE Signal Processing Letters, vol. 20, no. 4,
pp. 295–298, 2013.
[25] U. Vovk, F. Pernus, and B. Likar, “A review of methods for
correction of intensity inhomogeneity in MRI,” IEEE
Transactions on Medical Imaging, vol. 26, no. 3, pp. 405–421,
2007.
[26] C. Li, C. Xu, A. W. Anderson et al., “MRI tissue classification
and bias field estimation based on coherent local intensity
clustering: a unified energy minimization framework,” in
Proceedings of International Conference on Information Pro-
cessing in Medical Imaging, pp. 288–299, Springer-Verlag,
Williamsburg, VA, USA, July 2009.
[27] A. Buades, B. Coll, and J. M. Morel, “A non-local algorithm
for image denoising,” in Proceedings of IEEE Computer Society
Conference on Computer Vision & Pattern Recognition,
pp. 25-26, San Diego, CA, USA, July 2005.
[28] BrainWeb: Simulated Brain Database, 2004, http://brainweb.
bic.mni.mcgill.ca/brainweb/.
[29] C. Li, J. C. Gore, and C. Davatzikos, “Multiplicative intrinsic
component optimization (MICO) for MRI bias field esti-
mation and tissue segmentation,” Magnetic Resonance Im-
aging, vol. 32, no. 7, pp. 913–923, 2014.
[30] A. Keith and M. D. Johnson, The Whole Brain Atlas, http://
www.med.harvard.edu/aanlib/.
Hindawi
https://doi.org/10.1155/2019/2123079
Research Article
Comparing Strategies to Prevent Stroke and Ischemic Heart
Disease in the Tunisian Population: Markov Modeling Approach
Using a Comprehensive Sensitivity Analysis Algorithm
Olfa Saidi ,1 Martin O’Flaherty,2 Nada Zoghlami,1 Dhafer Malouche,1,3 Simon Capewell,2
Julia A. Critchley,4 Piotr Bandosz,2 Habiba Ben Romdhane,1 and Maria Guzman Castillo2
1
Cardiovascular Epidemiology and Prevention Research Laboratory, Faculty of Medicine of Tunis, University Tunis El Manar,
Tunis, Tunisia
2
Department of Public Health and Policy, University of Liverpool, Liverpool, UK
3
National Institute of Statistics and Data Analysis Tunis, Tunis, Tunisia
4
Population Health Research Institute, St George’s University of London, London, UK
Correspondence should be addressed to Olfa Saidi; olfa.saidi@yahoo.fr
Received 25 September 2018; Revised 27 November 2018; Accepted 18 December 2018; Published 29 January 2019
Copyright © 2019 Olfa Saidi et al. This is an open access article distributed under the Creative Commons Attribution License,
Background. Mathematical models offer the potential to analyze and compare the effectiveness of very different interventions to
prevent future cardiovascular disease. We developed a comprehensive Markov model to assess the impact of three interventions to
reduce ischemic heart diseases (IHD) and stroke deaths: (i) improved medical treatments in acute phase, (ii) secondary prevention
by increasing the uptake of statins, (iii) primary prevention using health promotion to reduce dietary salt consumption. Methods.
We developed and validated a Markov model for the Tunisian population aged 35–94 years old over a 20-year time horizon. We
compared the impact of specific treatments for stroke, lifestyle, and primary prevention on both IHD and stroke deaths. We then
undertook extensive sensitivity analyses using both a probabilistic multivariate approach and simple linear regression (meta-
modeling). Results. The model forecast a dramatic mortality rise, with 111,134 IHD and stroke deaths (95% CI 106567 to 115048)
predicted in 2025 in Tunisia. The salt reduction offered the potentially most powerful preventive intervention that might reduce
IHD and stroke deaths by 27% (−30240 [−30580 to −29900]) compared with 1% for medical strategies and 3% for secondary
prevention. The metamodeling highlighted that the initial development of a minor stroke substantially increased the subsequent
probability of a fatal stroke or IHD death. Conclusions. The primary prevention of cardiovascular disease via a reduction in dietary
salt consumption appeared much more effective than secondary or tertiary prevention approaches. Our simple but comprehensive
model offers a potentially attractive methodological approach that might now be extended and replicated in other contexts
and populations.
1. Background worldwide [4, 5]. They accounted for 15.2 million deaths
(15.0 million to 15.6 million) in 2015 [4]. According to the
Cardiovascular diseases (CVDs) cause nearly one-third of all World Health Organization (WHO), there are 15 million
deaths worldwide. 80% of these deaths occur in low-income people worldwide who suffer from stroke each year, among
and middle-income countries. Ischemic heart diseases them, 5 million die and another 5 million are left perma-
(IHD) and stroke account for the greatest proportion of nently disabled, causing a heavy burden for the family and
CVDs [1–3]. community.
The burden of IHD and stroke is considerable, and they The burden of stroke will increase significantly over the
are the first and second leading causes of death, respectively, next 20 years, particularly in developing countries [6]. Thus,
the analysis of the effectiveness of health promotion in- TP2,2

terventions is urgently required for appropriate planning to TP1,2
reduce this burden [7]. Healthy
TP1,1 population Sick
Traditional epidemiological study designs cannot ad- TP1,3 TP2,3
dress these issues; even clinical trials are usually restricted by
inclusion and exclusion criteria and not necessarily gener-
alizable to the entire population [8].
Mathematical modeling overcomes many of these lim- Dead
itations. It plays a crucial role in helping to guide the most
effective and cost-effective ways to achieve the goals of health
Figure 1: Markov diagram states and transition probabilities (each
promotion, designed and validated to guide health policies circle represents a Markov state and arrows indicate transition
and development strategies at several levels [9, 10]. probabilities).
Briefly, a model is a simplification of a real situation and
can encompass a simple, descriptive tool up to systems of
mathematical equations [11]. (ii) Cycle length: the length of time represented by a
The application of mathematical models in medicine has single stage (or cycle) in a Markov process: Markov
proved useful and has become more frequent, especially for models are developed to simulate both short-cycle
cardiovascular diseases (CVDs) [12–16] and assessing po- and long-term processes (e.g., cardiovascular
tential impacts of policies or interventions designed to alter diseases).
disease trajectories in Tunisia [17, 18]. (iii) Transition probabilities: the matrix of probabilities
A commonly used technique is Markov modeling, an ap- of moving between health states from one stage to
proach that models groups of individuals transitioning across the next.
specified pathways, informed by transition probabilities [18–20].
Due to the frequent uncertainty of the Markov model’s
inputs, sensitivity analyses to assess the robustness of the 2.2. Process of Mathematical Modeling Based on Markov
model results are strongly recommended by modeling Approach. Before starting the data collection and the cal-
guidelines. Such uncertainty analyses assess confidence in a culation, we first defined our model by specifying the dif-
chosen course of action and ascertain the value of collecting ferent states that can be included based on the literature and
additional information to better inform the decision [21]. expert opinions:
Our study aims (i) to describe a comprehensive Markov
S � 􏼈S1 , S2 , . . . , S3 􏼉, set of states in the process. (1)
model based on both a probabilistic multivariate approach
and simple linear regression metamodeling and (ii) using the Having specified the structure of the model in terms of
model to evaluate the effects of increases in uptake of stroke the possible transitions between states, we defined the
treatments lifestyle changes and primary prevention among transition probabilities based on available data.
the Tunisian population aged 35–94 years old in 2025. We The transition probability (TPij ) is defined as a condi-
examined three interventions: (a) improved medical treat- tional probability (Pt (sj /si )) of making a transition (mov-
ments in the acute phase, (b) secondary prevention of stroke ing) from state i to state j during a single cycle (t).
by increasing the prescribing of statins, and (c) primary Additionally, transition probabilities are stratified by sex and
prevention aiming to reduce salt intake. age groups ag ∈ 􏽮c1 , c2 , . . . , cg 􏽯, where c1 , c2 , . . . , cg repre-
sents a set of age groups [19, 22].
2. Methods One of the goals of the Markov model is to study the
potential effects of some health promotion interventions.
In this study, we describe the development of a Markov IHD We modeled first a baseline scenario and then the in-
and stroke model. tervention scenarios.
For the baseline scenario, we assumed no change will
2.1. Definition of Markov Model. Markov models are a type happen during the period “T” of study (20 years) in either
mathematical model based on matrix algebra which de- the present uptake rates of medical therapies or population
scribes the transitions a cohort of patients make among a level uptakes of specific nutrients.
number of mutually exclusive and exhaustive health states Based on the model structure in Figure 1, we assume that
during a series of short intervals or cycles. In this model, a we will apply three interventions to study their impact on
patient is always in one of the finite number of health states; mortality in the future (over the 20 years: from t � year 1 to
events are modeled as transitions from one state to another year 20).
and contribution of utility to overall prognosis depend on We define a policy Π � (I0 , I1 , I2 , I3 ) as a set of the
length of time spent in health states [19]. The components of health promotion interventions.
a Markov model are shown below (Figure 1): I0 refers the baseline scenario: we assumed no change
(i) States: the set of distinct health states under con- will happen during 20 years.
sideration in the model, together with the possible I1 : scenario aimed to improve medical treatments in
transitions between them. acute phase.
I2 : scenario for secondary prevention. Step III. Calculate the number of people in each state for the
I3 : scenario for primary prevention. interventions scenarios
For the interventions scenarios, the model required a
Starting by the baseline scenario, the process of the base estimate of risk reduction in deaths and the age effect to
Markov model is based on the two first steps below: calculate policy effectiveness (ΠIie ) of each intervention.
The policy effectiveness is defined by the following
Step I. Calculate probabilistic transition probabilities formula:
Probabilistic sensitivity analysis aims to fully evaluate the Ii
combination of uncertainty in all model inputs (including 􏽙 � 1 − RRi × ae 􏼁, (4)
transition probabilities) simultaneously on the robustness of e
model results.
The point estimates in the model can be replaced with where RRi is the risk reduction associated with the in-
probability distributions, where the mean of the distribution tervention Ii(i�1:3) obtained from previous randomized
reflects the best estimate of the parameter. controlled trials and meta-analyses to estimate the re-
In this step, each input parameter (TPij ) is assigned an duction in age-specific deaths and ae represents the age
appropriate statistical distribution, and a Monte Carlo effects for each intervention risk reduction value. ΠIie is then
simulation is run multiple times (e.g., 1000). The iteration is used to scale the transition probabilities connected to death
stopped when the difference of the outcomes is sufficiently states.
small. In this stage, after recalculating the TPs, we define the
We assume for example that TPs follow the distribution number of people in each state (healthy people (NHPIi ), sick
defined by beta (α, β), where α probability distributions are (NSIi ), and deaths (NDIi ) for all sex and age groups from
defined on the interval [0, 1] parameterized by two positive t � year 1 to year T for the interventions scenarios Ii , given
shape parameters, denoted by α and β, that appear as ex- by the following equations:
ponents of the random variable and control the shape of the Ii
distribution [23]. ⎝1 − TP − TP × 􏽙⎞
NHPlit � NHPlit−1 × ⎛ ⎠,
1,2 1,3
e
Step II. Calculate the number of people in each state for the NSlit � NSlit−1 + 􏼐NHPlit−1 × TP1,2 􏼑,
baseline scenario Ii
The next step is to define the number of people in each NDlit � NDlit−1 + 􏼐NHPlit−1 × TP1,3 􏼑 + ⎛
⎝NSli × TP × 􏽙⎞
⎠,
t−1 2,3
state based on the TPs between states and the number of e
people in the starting state “healthy people.”
In this stage, we define the number of people in each state LYlit � NHPlit − NDlit .
for the baseline scenario (healthy people (NHP), sick (NS), (5)
and deaths (ND)) for all gender and age groups from t � year
1 to year T, given by the following equations:
NHPt � NHPt−1 × 􏼐1 − 􏽘 TPi,j 􏼑, Step IV. Calculate the final outputs (DPPs and LY)
Finally, we calculate the total number of CVD deaths
NSt � NSt−1 + 􏼐NHPt−1 × TP1,2 􏼑, (ischemic stroke and IHD deaths) that could be prevented or
postponed (DPPs) and the life years gained (LY) under each
NDt � NDt−1 + 􏼐NHPt−1 × TP1,3 􏼑 + 􏼐NSt−1 × TP2,3 􏼑. specific scenario (equations (6) and (7)), compared to the
(2) baseline scenario for all sex and age groups from t � year 1 to
year T for the interventions scenarios Ii .
Another indicator can be calculated based on this model,
that is, life years gained (LY) defined by the following DPPlist � NDlit − NDl0
t , (6)
equation:
LYligt � LYlit − LYl0
t . (7)
LYt � NHPt − NDt . (3)
The model is then run several times (e.g., 1000 simu-

lations). For each simulation model, parameter values are Step V. Linear regression metamodeling
randomly drawn from each of the distributions, and the Most current modeling studies are limited to the first
expected model outcome is recorded. The 1000 simulations four stages and ignore this important step of metamodeling
result in a distribution of expected model outcomes to analyze which model inputs are most influential in af-
(e.g., deaths), which reflects the overall parameter un- fecting the results. The goal of metamodeling was thus to
certainty in the mode [24]. increase the transparency of decision-making analytic
At the end of the simulation, the mean as well as the models and better communicate their results.
lower bound (LB) and upper bound (UB) of 95% confidence This step is based on the application of a simple linear
interval of the inputs and outputs will be calculated, which regression metamodel (LRM) for the optimal
correspond to Steps I and II presented in Algorithm 1. policy (Figure 2) [25].
Step I. Calculate probabilistic transition probabilities

for all k ∈ {1 Male, 2 Female} do
for all age groups ag ∈ 􏽮c1 , c2 , . . . , cg 􏽯 do
for t � 1.....T do
TPi,j ⟵ beta(αi,j , βi,j )
end for
end for
end for
Step II. Calculate the number of people in each state NHP, NS, and ND (Baseline scenario)
for t � 1.....T do
NHPt � NHPt−1 × (1 − 􏽐 TPi,j )
NSt � NSt−1 + (NHPt−1 × TP1,2 )
NDt � NDt−1 + (NHPt−1 × TP1,3 ) + (NSt−1 × TP2,3 )
LYt � NHPt + NDt
end for
end for
end for
Step III. Calculate the number of people in each state for the interventions scenarios
for all intervention Ii ∈ Π � (I1 , . . . , I3 ) do
ΠIie � 1 − (RRi × ae ) where RRIi is the intervention risk reduction value and ae its age effects
for t � 1 . . . T
NHPlit � NHPlit−1 × (1 − TP1,2 − TP1,3 × ΠIie )
NSlit � NSlit−1 + (NHPlit−1 × TP1,2 )
NDlit � NDlit−1 + (NHPlit−1 × TP1,3 ) + (NSlit−1 × TP2,3 × ΠIie )
LYlit � NHPlit − NDlit
end for
end for
end for
end for
Step IV. Calculate the final outputs (DPPs and LY)
for all intervention Ii ∈ Π � (I1 , . . . , I3 ) do
DPPlist � 􏽐T1 NDlit − NDl0 t
LYligt � 􏽐T1 LYlit − LYl0
t end for
end for
end for
end for
Step V. Linear regression metamodeling
Calculate the βi,j coefficients of the parameters
Solve the equation ND � β0 + βi,j TPi,j + ε
end for
end for
ALGORITHM 1: Comprehensive algorithm of the Markov model and of sensitivity analysis.
The original motivation for metamodeling was to define other coefficients, are interpreted relative to a 1-unit change in
a simpler mathematical relationship between model outputs each parameter on the original scales; and ε is the residual term.
and inputs than the actual model. Furthermore, the absolute value of the coefficients of the
The LRM is defined by the following formula: parameters “βi,j ” can be used to rank parameters by their
ND � β0 + βi,j TPi,j + ε, (8) importance: the higher the coefficient is, the more the
variable is important and relevant.
where ND is the number of deaths (output of the model); β0 The algorithm below summarizes the process with three
(the intercept) is the expected outcome when all parameters are states that could be generalized to n states depending on the
set to zero; TPi,j , the transition probabilities (inputs); βi,j , the context to study (Algorithm 1).
Outputs (e.g. pressure control, glucose control, lipid lowering, salt uptake,
Real life Inputs (TPij)
deaths (ND)) and smoking cessation).
2.6. Prevention Scenarios. In this paper we evaluated the

following three scenarios:
(1) I1 : the first scenario aimed to improve medical
LRM:
treatments in the acute phase: increasing throm-
ND = β0 + βij TPij + ε
bolysis prescribing from 0% to 50% and hospitali-
zation in stroke units from 10% to 100% in Tunisia.
Figure 2: Simple linear regression metamodel (LRM) to sum- (2) I2 : the second scenario aimed to act on medical
marize the relationship between model inputs and outputs.
treatments after a stroke: increasing the prescribing
of statins from 11% to 100% (secondary prevention).
2.3. Case Study. The algorithm introduced above has been (3) I3 : the third scenario aimed to reduce the con-
implemented using R (Version R.3.2.0.) software and ap- sumption of salt by 30%, from 14 grams to 9 grams
plied on Tunisian data. The source codes are available from per day (primary prevention).
the authors upon request.
Our model predicts both IHD and stroke deaths in 2025 Total policy refers to the combined effects of all the three
among the Tunisian population aged 35–94 years old, both previous strategies: acute treatment + secondary pre-
men and women, and compares the impact of specific vention + primary prevention.
treatments for stroke, lifestyle changes, and primary
prevention. 2.7. Modeling Policy Effectiveness and Its Impact in Mortality.
The model applies the relative risk reduction (RRR), as
mentioned in the algorithm above, quantified for each in-
2.4. Model Structure. Data were integrated and analyzed tervention scenario in previous randomized controlled trials
using a closed cohort model based on a Markov approach, and meta-analyses based on international studies (data are
with transition probabilities starting from population free of detailed in the Appendix in the supplementary materials
ischemic stroke and moving to health states reflecting the (Table 8)).
natural history of ischemic stroke [26, 27]. Therefore, the
starting states are defined by the size of the population and
the number of strokes occurring within this population. The 2.8. Data Sources. Published and unpublished data were
number of persons moving from the starting states to the identified by extensive searches, complemented with specif-
stroke and death states is estimated by the transition ically designed surveys. Data items included (i) number of
probabilities (Table 9 in the appendix in the supplementary stroke patients (minor and major), (ii) uptake of specific
materials). medical and surgical treatments, (iii) population data in the
There are two absorbing states: IHD and stroke deaths initial study year (2005) and (iv) mortality data (after one year
and non-IHD and stroke deaths as competing risks for (data in 2006) and after 5 years (data in 2010)). Data sources
mortality. IHD and stroke have several risk factors in are detailed in Appendixin the supplementary materials.
common; increased salt intake is associated with hyper-
tension which is one of the major risk factors for both stroke 2.9. Model Outputs. The outputs of this model are the
and IHD. Therefore, any policy aimed at decreasing pop- prediction of stroke and IHD deaths prevented or postponed
ulation level’s intake of salt will reduce the risk of both (DPPs) and the life years gained (LY) among the Tunisian
diseases [28]. population aged 35–94 years old starting from 2005 over a
Potential overlaps between the healthy, minor, major twenty-year time period(to 2025) with and without possible
stroke, and deaths are managed by calculating the condi- interventions to reduce this mortality.
tional probabilities of membership. A key element of the We modeled all the intervention scenarios to calculate
model is the calculation of the transition probabilities (TPs), the total number of CVD deaths (ischemic stroke and
particularly those related to stroke and IHD mortality IHD deaths) that may be prevented or postponed and the
(Figure 3). The TP calculations are detailed in Table 10 in the LY for each specific scenario compared to the baseline
Appendix in the supplementary materials. scenario.
3. Results
2.5. Baseline Scenario. In the baseline scenario, we assumed
there would be no change over the 20-year model period in 3.1. Baseline Scenario. In the baseline scenario, the model
the present uptake rates of acute phase medical treatments forecast 111140 [95% CI 106570 to 115050] IHD and is-
(thrombolysis, aspirin, and stroke unit) or treatments for chemic stroke deaths for people aged 35–94 years between
secondary prevention following stroke (aspirin, statin, an- 2005 and 2025, including 68890 [95% CI 65730 to 72350]
ticoagulants, blood pressure control, and smoking cessation) among men and 42250 [95% CI 38840 to 44600] among
or treatment and policies for primary prevention (blood women.
TP4,4
Minor
TP2,4 stroke and
TIA
TP4,5
Minor
stroke and TP2,5
TP1,1 TIA Nonstroke
TP1,2
and IHD
TP1,5 deaths
Stroke-free
TP2,3 TP2,6 TP3,5
population
TP1,6 TP1,6
TP3,3 TP4,6
TP1,3 Stroke and
IHD
Major deaths
stroke
TP3,6
Figure 3: Stroke model structure. TIA: transient ischemic attack; IHD: ischemic heart diseases.
The model estimated that the acute stroke treatment and postponed (DPPs) by the salt reduction intervention. The R2
secondary prevention following stroke would, respectively, is 0.8999, suggesting that 89. 99% of the variance in the
prevent 230 [95% CI 200 to 260] deaths due to stroke and model outcomes is explained by our model (Table 3).
IHD and 1920 [95% CI 1830 to 2000] in 2025, whilst primary Figure 5 shows the five first important parameters based
prevention could prevent 20330 [95% CI 20050 to 20610] on the absolute value of the coefficients of the parameters. In
cumulative deaths due to stroke and IHD in 2025. our case study, the uncertainty of the probability of minor
In terms of life years, 150 [95% CI 130 to 180] and 2390 stroke in the first year has the greatest impact on the stroke
[95% CI 2300 to 2490] would be gained in 2025 by acute and IHD deaths estimates associated with salt reduction,
stroke treatment and secondary prevention, respectively, followed by the probability of the stroke-free population to
whereas for primary prevention (blood pressure control, die from stroke and IHD causes in the year 1 and the
glucose control, lipid lowering, and smoking cessation), probability of recurrent stroke in ischemic stroke patients
14590 [95% CI 14350 to 14820] life years would be gained in after one year. Although the main objective of the meta-
2025 (Table 1). regression is to identify the most important parameters of
the model, it could also serve to give a rough idea of the size
3.2. Scenario Projections of the effect of each parameter. We know that for each unit
increase in the independent variable, our outcome should
2
3.2.1. Improvement on Acute Stroke Treatment. If we adopt a increase by 􏽢I units. For example, the probability for the
strategy to increase thrombolysis uptake from 0% to 50% stroke-free population to have first stroke in the year 1 has
and stroke units use from 10% to 100%, 600 [95% CI 550 to an absolute coefficient of 646.35; this means that for each
650] fewer deaths could be achieved by 2025 (Table 2). 10% risk increase in this probability, strokes and IHD will
In terms of life years, 370 [95% CI 330 to 410] would be increase by 65 IHD and stroke deaths. However, this should
gained in 2025 by increasing thrombolysis uptake (Figure 4). be interpreted with caution since the meta-analysis re-
gression model assumes a linear relationship between
3.2.2. Improvement on Secondary Prevention. If the uptake outcomes and inputs which is not the case in a Markov
of statins for secondary prevention following stroke could be model (Figure 5).
increased from 11% to 100%, 3300 [95% CI 3190 to 3410]
fewer deaths could be avoided by 2025 (Table 2).
In terms of life years, 4120 [95% CI 4000 to 4250] would 4. Discussion
be gained in 2025 (Figure 4). In this study, we have developed a simple but comprehensive
Markov model and used it to identify key factors that predict
mortality from stroke and IHD in Tunisia in the future, as
3.2.3. Food Policies. If the Tunisian government imple-
well as the potential impacts of some medical and health
mented a strategy recommended by the WHO to reduce the
policies.
daily consumption of salt by 30 % (from 14 g/day to 9 g/day),
Different mathematical models have been highly used in
30240 [95% CI 29900 to 30580] deaths could be avoided by
medical decision making [19, 29–32], but the technique of
this scenario in 2025 (Table 2). This results in 20630 [95% CI
metamodeling is less developed in medicine. It has been used
20350 to 20910] life year gain by 2025 (Figure 4).
to identify the importance of variables that can justify the
best medical decisions among pregnant women with deep
3.3. Linear Regression Metamodeling of the Optimal Policy. vein thrombosis [33]. Additionally, linear regression met-
The table below shows the results of linear regressing pa- amodels have also been used in some epidemiological
rameters on the stroke and IHD deaths prevented or studies [25, 34].
Table 1: Life years and deaths due to stroke and IHD estimations in 2025 keeping the same practices of 2005 by gender.
Life years [95% CI] Stroke and IHD deaths [95% CI]
Men
Acute stroke treatment 80 [70 to 100] −140 [−170 to −120]
Secondary prevention following stroke 1500 [1420 to 1580] −1170 [−1240 to −1110]
Primary prevention 12180 [11960 to 12400] −16760 [−17020 to −16510]
Policy total∗ 6830 [6700 to 6990] −8530 [−8710 to −8340]
Women
Both
∗
Total policy refers to the combined effects of all the three previous strategies: acute treatment + secondary prevention + primary prevention.
Table 2: Life years and deaths due to stroke and IHD by in- assumptions with clear justification (see Appendix in the
corporating strategies in 2025 by gender. supplementary materials).
Stroke and IHD deaths [95%
Another advantage characterizing our approach and
CI] differentiating it from the univariate traditional method of
Men
deterministic sensitivity analysis is that it allows varying all
Acute stroke treatment −350 [−390 to −310] the parameters of the model simultaneously and thus ex-
Secondary prevention following ploring the whole parameter space. The use of only one-way
−2060 [−2150 to −1970] sensitivity analyses is limited compared with multiway
stroke
Primary prevention −24500 [−24810 to −24200] analyses [35].
Policy total −23940 [−24240 to −23640] Furthermore, to increase the transparency of decision-
Women making analytic models, we have used metamodeling in this
Acute stroke treatment −220 [−250 to −190] study by applying a simpler mathematical relationship be-
Secondary prevention following tween model outputs and inputs to analyze which are the
−1240 [−1310 to −1170]
stroke most influential inputs on the output.
Primary prevention −10630 [−10830 to −10430]
Our metamodeling approach summarizes the results of
Policy total −17050 [−17300 to −16800]
Both
the model studied in a transparent way and reveals its
Acute stroke treatment −600 [−650 to −550] important characteristics.
Secondary prevention following The intercept of the regression result is the expected
−3300 [−3410 to −3190] outcome when all parameters are set to zero. The other
stroke
Primary prevention −30240 [−30580 to −29900] coefficients are interpreted relative to a 1-unit change in each
Policy total∗ −40990 [−41390 to −40600] parameter on the original scales. For example, in our case,
∗
Total policy refers to the combined effects of all the three previous changing the risk of the first minor stroke in the first year
strategies: acute treatment + secondary prevention + primary prevention. from the actual value to 1 increases the number of stroke and
IHD deaths by 646 deaths. In addition, the regression co-
efficients describe the relative importance of the uncertainty
in each parameter. And the use of the linear metamodeling
This analysis is the first modeling study of stroke and regression method can overcome the limitations of other
IHD mortality in Tunisia, based on a rigorous and com- traditional statistical methods [36].
prehensive modeling approach consisting of three stages: In addition, models often ignore the interaction effect
(i) Apply the Markov model in medical and strategic between the input parameters of a model when defining the
decision making. results. However, in our study, the use of conditional
(ii) Provide a probabilistic sensitivity analysis. probabilities allows us to take into account interactions in
our algorithm.
(iii) Use a linear regression metamodeling for the op- We also analyzed, in parallel of the deaths prevented or
timal policy. postponed (DPPs), the life years gained according to the
This model has several strengths. First it requires mul- different scenarios. Thus, our approach allows interpreting
tiple epidemiological national data on ischemic stroke and two key indicators in epidemiology via a rigorous and
demographics. In general, the data used in the model were of comprehensive mathematical algorithm.
good quality. However, some assumptions were necessary to Nevertheless this modeling approach has also some
fill gaps in the missing information. We made transparent limitations; in fact, a major limitation of our work is that the
15000
10000
Life years
5000
Acute stroke Secondary prevention Primary Policy total

treatment following stroke prevention
Strategies
Women
Men
Both
Figure 4: Life years estimations by incorporating strategies by gender in 2025.
Table 3: Regression coefficients from metamodeling on the stroke and IHD deaths by the salt reduction intervention.
Parameters Dependent variable (Y): stroke and IHD deaths
Intercept 1337.86
TP1,2 : probability for the stroke-free population to
646.35
have first stroke in the year 1
TP1,3 : probability for the stroke-free population to
6.18
have first major stroke in the year 1
TP1,6 : probability for the stroke-free population to die
38.49
from stroke and IHD causes in the year 1
TP1,1 : probability for population free of stroke 2.45
TP2,3 : probability of recurrent stroke in ischemic
13.05
stroke patients after 1 year
TP2,6 : probability for the minor stroke patients to die
6.67
from stroke and IHD causes in the year 1
TP2,4 : probability for first minor stroke (1st year) to
0.45
minor stroke subsequent years
TP3,3 : probability of recurrent stroke in ischemic
−6.49
stroke patients after 5 years
TP4,6 : probability for the stroke patients to die from
stroke and IHD deaths causes 1 year after first −1.72
admission
TP4,4 : probability for minor stroke subsequent years −2.96
TP3,6 : probability for the major stroke patients to die
−0.10
from stroke and IHD causes
Observations 1000
R2 0.8999
model is a closed cohort, and demographic changes were not Another limitation of our study is not to consider in the
considered. model the costs of the strategies.
Linear regression is the metamodeling approach most In terms of public health, the present modeling study
widely used by many researchers for its simplicity and ease of focuses on the future impact of ischemic stroke treatment
use. Thus, our linear approximation of all input parameters scenarios and population-level policy interventions on
can be considered as a limit, as state-transition models are ischemic stroke and IHD deaths and life years gained in
nonlinear in general. In all metamodeling approaches, the Tunisia. The model forecast a dramatic rise in the total
loss of some information is always inevitable [25]. cumulative number of IHD and Stroke deaths: by 2025,
700 646 in Tunisia over 2005–2025. This Large increase in stroke and
600
IHD mortality in Tunisia needs many actions not only in
Stroke and IHD deaths
acute stroke treatment such as implementing more basic and

500
organized stroke units but also in population level primary
400 prevention such as salt reduction in order to manage and
300 treat acute strokes and to alleviate the global burden of these
200 diseases.
100 38 13 7 6
Our study highlights the powerful impact of salt re-
0
duction on deaths from stroke and IHD. Furthermore, the
TP1,2 TP1,6 TP2,3 TP2,6 TP3,3 reducing dietary salt intake across the population appears an
Transition probabilities effective way of reducing heart disease events and saving
substantial costs. This result matches with that of the
Figure 5: Five first important parameters in the model.
mathematical model. Indeed our metamodeling highlights
that the probability of the first minor stroke among the
this number was estimated to reach more than 100,000 healthy population has the greatest impact on the stroke and
deaths. IHD death estimations, which confirms the importance of
Secondly, the model shows that the rise of thrombolytic primary prevention. Prevention is thus the best strategy to
treatment and hospitalization in intensive care units of fight against stroke and IHD deaths.
stroke increased statin use for secondary prevention pale in
comparison with the salt reduction impact on future deaths Data Availability
(1%, 3% vs 27% deaths prevented by 2025).
The benefits of thrombolytic treatment among patients Data used are presented in Appendix in the supplementary
with acute ischemic stroke are still matter of debate: materials.
thrombolytic treatment increases short-term mortality and
symptomatic or fatal intracranial haemorrhage but decreases Disclosure
longer term death or dependence [25]. Many studies proved
that stroke units have significant benefit to patient outcomes The results of the manuscript were presented in the 62nd
in terms of reducing mortality, morbidity, and improving Annual Scientific Meeting, Society for Social Medicine,
functional independence. Stroke unit care was also cost University of Strathclyde, Glasgow, on the 5th and 7th of
effective [37–39]. September 2018. The funders had no role in study design,
The substantial effect of salt reduction intervention data collection and analysis, decision to publish, or prepa-
found in our study is consistent with the literature [40, 41]. ration of the manuscript. The MedCHAMPS team had ac-
High salt intake is associated with significantly increased cess to all data sources and has the responsibility for the
risks of stroke and total cardiovascular disease ranging from contents of the report. On behalf of EC, FP7 funded
a 14% to 51% increase, depending on salt intake and pop- MedCHAMPS project (MEDiterranean studies of Cardio-
ulations [42–44]. vascular Disease and Hyperglycaemia: Analytical Modeling
Furthermore, our results are consistent with a modeling of Population Socio-economic transitions).
study in Tunisia using a different approach. Different
strategies for salt reduction were associated with substantial Conflicts of Interest
lowering of CVD mortality and were cost saving apart from
The authors declare that they have no conflicts of interest.
health promotion [45] and consistent with the Rose hy-
pothesis that the population level strategies are more ap-
propriate in terms of less medicalization [46].
Authors’ Contributions
MOF, OS, SC, and JC conceived the idea of the study. OS,
5. Conclusions DM, and MOF led the project supervised by HBR, JC, and
SC. OS, MG, NZ, and PB. assembled the datasets, extracted
The approach presented here is attractive since it is based on the data, and populated the models with input. OS, NZ, DM,
a simple comprehensive algorithm to present the results of MG, and MOF wrote the first draft of the paper, and OS,
sensitivity analysis from the Markov model using linear HBR, MOF, SC, and JC finalized the manuscript. All authors
regression metamodeling. This approach can reveal im- contributed to the analysis, intellectual content, critical
portant characteristics of Markov decision process including revisions to the drafts of the paper and approved the final
the base-case results, relative parameter importance, inversion. HBR is the guarantor.
teraction, and sensitivity analyses.
Thus, we recommend using this algorithm for Markov Acknowledgments
decision process; it can be the object of creation of a
complete modeling package in R software and can be ex- We thank the MedCHAMPS and RESCAPMED advisory
tended to other contexts and populations. group for comments on study design and emerging findings.
In terms of public health, we forecast that absolute Particular thanks go to Pat Barker for managing the overall
numbers of IHD and stroke deaths will increase dramatically project. We thank also the Tunisian team especially Dr Afef
Skhiri and Dr Olfa Lassoued who participated in data col- [13] M. Ortegon, S. Lim, D. Chisholm, and S. Mendis, “Cost ef-
lection and Professor Faycel Hentati for his opinions and fectiveness of strategies to combat cardiovascular disease,
comments on the work. We thank Professor K. Bennett for diabetes, and tobacco use in sub-Saharan Africa and South
her comments on this paper. This study was funded by the East Asia: mathematical modelling study,” BMJ, vol. 344,
European Community’s Seventh Framework Program (FP7/ no. 1, p. e607, 2012.
[14] L. C. Leea, G. S. Kassabb, and J. M. Guccionec, “Mathematical
20072013) under grant agreement n223075, the Med-
modeling of cardiac growth and remodeling,” Wiley In-
CHAMPS project. MOF was also partially supported by the terdisciplinary Reviews: Systems Biology and Medicine, vol. 8,
UK MRC. JC and SC are supported by the UK Higher no. 3, pp. 211–226, 2017.
Education Funding Council. [15] G. Elena, “Computational and mathematical methods in
cardiovascular diseases,” Computational and Mathematical
Methods in Medicine, vol. 2017, Article ID 4205735, 2 pages,
Supplementary Materials 2017.
The supplementary materials submitted contain sources of [16] S. Basu, David Stuckler, S. Vellakkal, and E. Shah, “Dietary salt
reduction and cardiovascular disease rates in India: a
data and all data used in the model. (Supplementary
mathematical model,” PLoS One, vol. 7, no. 9, 2012.
Materials) [17] O. Saidi, N. Ben Mansour, M. O’Flaherty, S. Capewell,
J. A. Critchley, and H. B. Romdhane, “Analyzing recent
References coronary heart disease mortality trends in Tunisia between
1997 and 2009,” PLoS One, vol. 8, no. 5, Article ID e63202,
[1] N. D. Wong, “Epidemiological studies of CHD and the 2013.
evolution of preventive cardiology,” Nature Reviews Cardi- [18] O. Saidi, M. O’Flaherty, N. Ben Mansour et al., “Forecasting
ology, vol. 11, no. 5, pp. 276–289, 2014. Tunisian type 2 diabetes prevalence to 2027: validation of a
[2] S. Ebrahim and G. D. Smith, “Exporting failure? Coronary simple model,” BMC Public Health, vol. 15, no. 1, p. 104, 2015.
heart disease and stroke in developing countries,” In- [19] F. A. Sonnenberg and J. R. Beck, “Markov models in medical
ternational Journal of Epidemiology, vol. 30, no. 2, pp. 201–5, decision making,” Medical Decision Making, vol. 13, no. 4,
2001. pp. 322–338, 2016.
[3] T. A. Gaziano, A. Bitton, S. Anand et al., “Growing epidemic [20] K. Szmen, B. Unal, S. Capewell, J. Critchley, and
of coronary heart disease in low- and middle-income coun- M. O’Flaherty, “Estimating diabetes prevalence in Turkey in
tries,” Current Problems in Cardiology, vol. 35, no. 2, 2025 with and without possible interventions to reduce
pp. 72–115, 2010. obesity and smoking prevalence, using a modeling approach,”
[4] GBD 2015. Mortality and Causes of Death Collaborators, International Journal of Public Health, vol. 60, no. 1, pp. 13–21,
“Global, regional, and national life expectancy, all-cause 2015.
mortality, and cause-specific mortality for 249 causes of [21] A. H. Briggs, M. C. Weinstein, E. A. L. Fenwick, J. Karnon,
death, 1980–2015: a systematic analysis for the Global Burden M. J. Sculpher, and A. D. Paltiel, “Model parameter estimation
of Disease Study 2015,” The Lancet, vol. 388, no. 10053, and uncertainty: a report of the ISPOR-SMDM modeling
pp. 1459–1544, 2016. good research practices task force-6,” Value in Health, vol. 15,
[5] G. J. Hankey, “The global and regional burden of stroke,” The no. 6, pp. 835–842, 2012.
Lancet Global Health, vol. 1, no. 5, pp. e239–e240, 2013. [22] M. L. Puterman, Markov Decision Processes: Discrete Sto-
[6] WHO, “Global burden of stroke,” 2015, http://www.who.int/ chastic Dynamic Programming, Wiley, New York, NY, USA,
cardiovascular_diseases/en/cvd_atlas_15_burden_stroke.pdf? 2005.
ua�1. [23] A. Mahanta, “Beta distribution. Kaliabor college,” http://
[7] World Health Organization, “Health promotion,” 2018, www.kaliaborcollege.org/pdf/Mathematical_Excursion_to_
http://www.who.int/topics/health_promotion/en/. Beta_Distribution.pdf.
[8] P. J. Easterbrook, M. C. Doherty, J. H. Perrie, J. L. Barcarolo, [24] K. Kuntz, F. Sainfort, M. Butler et al., “Decision and simu-
and G. O. Hirnschall, “The role of mathematical modelling in lation modeling in systematic reviews,” Methods Research
the development of recommendations in the 2013 WHO Report 11(13)-EHC037-EF, AHRQ Publication, Rockville,
consolidated antiretroviral therapy guidelines,” AIDS, vol. 28, MA, USA, 2013.
pp. S85–S92, 2014. [25] H. Jalal, D. Bryan, F. Sainfort, M. Karen, and Kuntz, “Linear
[9] C. J. E. Metcalf, W. J. Edmunds, and J. Lessler, “Six challenges regression metamodeling asa tool to summarize and present
in modelling for public health policy,” Epidemics, vol. 10, simulation model results,” Medical Decision Making, vol. 33,
p. 9396, 2015. no. 7, pp. 880–890, 2013.
[10] M. C. Weinstein, P. G. Coxson, L. W. Williams, T. M. Pass, [26] V. Kapetanakis, F. E. Matthews, and A. van den Hout, “A
W. B. Stason, and L. Goldman, “Forecasting coronary heart semi-Markov model for stroke with piecewise-constant
disease incidence, mortality, and cost: the Coronary Heart hazards in the presence of left, right and interval censor-
Disease Policy Model,” American Journal of Public Health, ing,” Statistics in Medicine, vol. 32, no. 4, pp. 697–713, 2012.
vol. 77, no. 11, pp. 1417–1426, 1987. [27] C. Eulenburg, S. Mahner, L. Woelber, and K. Wegscheider, “A
[11] J. Critchley and S. Capewell, “Why model coronary heart systematic model specification procedure for an illness-death
disease?,” European Heart Journal, vol. 23, no. 2, pp. 110–116, model without recovery,” PLoS One, vol. 10, no. 4, Article ID
2002. e0123489, 2015.
[12] G. R. P. Garnett, S. Cousens, T. B. Hallett, R. Steketee, and [28] F. J. He and G. A. MacGregor, “Role of salt intake in pre-
N. Walker, “Mathematical models in the evaluation of health vention of cardiovascular disease: controversies and chal-
programmes,” The Lancet, vol. 378, no. 9790, pp. 515–525, lenges,” Nature Reviews Cardiology, vol. 15, no. 6, pp. 371–377,
2011. 2018.
[29] J. F. Moxnes and Y. V. Sandbakk, “Mathematical modelling of [46] G. Rose, “Strategy of prevention: lessons from cardiovascular
the oxygen uptake kinetics during whole-body endurance disease,” BMJ, vol. 282, no. 6279, pp. 1847–1851, 1981.
exercise and recovery,” Mathematical and Computer Mod-
elling of Dynamical Systems, vol. 24, no. 1, pp. 76–86, 2018.
[30] R. B Chambers, “The role of mathematical modeling in
medical research: research without patients?,” Ochsner
Journal, vol. 2, no. 4, pp. 218–223, 2000.
[31] F. Achana, J. Alex, D. K. Sutton et al., “A decision analytic
model to investigate the cost-effectiveness of poisoning
prevention practices in households with young children,”
BMC Public Health, vol. 16, no. 1, p. 705, 2016.
[32] O. Alagoz, H. Hsu, A. J. Schaefer, and M. S. Roberts, “Markov
decision processes: a tool for sequential decision making
under uncertainty,” Medical Decision Making, vol. 30, no. 4,
pp. 474–483, 2010.
[33] J. F. Merz and M. J. Small, “Measuring decision sensitivity: a
combined monte carlo-logistic regression approach,” Medical
Decision Making, vol. 12, no. 3, pp. 189–196, 1992.
[34] P. M. Reis dos Santos and M. Isabel Reis dos Santos, “Using
subsystem linear regression metamodels in stochastic simu-
lation,” European Journal of Operational Research, vol. 196,
no. 3, pp. 1031–1040, 2008.
[35] A. Moran, D. Gu, D. Zhao et al., “Future cardiovascular
disease in China: Markov model and risk factor scenario
projections from the Coronary Heart Disease Policy Model-
China,” Circulation: Cardiovascular Quality and Outcomes,
vol. 3, no. 3, pp. 243–252, 2010.
[36] D. Coyle, M. J. Buxton, and B. J. OBrien, “Measures of im-
portance for economic analysis based on decision modeling,”
Journal of Clinical Epidemiology, vol. 56, no. 10, pp. 989–997,
2003.
[37] K. Meisel and B. Silver, “The importance of stroke units,”
Medicine and Health, Rhode Island, vol. 94, no. 12,
pp. 376-377, 2011.
[38] P. Langhorne, “Collaborative systematic review of the rand-
omised trials of organised inpatient (stroke unit) care after
stroke,” BMJ, vol. 314, no. 7088, pp. 1151–1159, 1997.
[39] B. J. T. Ao, P. M. Brown, V. L. Feigin, and C. S. Anderson,
“Are stroke units cost effective? Evidence from a New Zealand
stroke incidence and population-based study,” International
Journal of Stroke, vol. 7, no. 8, pp. 623–630, 2011.
[40] N. Bruce, The Effectiveness and Costs of Population In-
terventions to Reduce Salt Consumption, WHO Library
Cataloguing-in-Publication Data, Geneva, Switzerland, 2006,
http://www.who.int/dietphysicalactivity/Neal_saltpaper_2006.
pdf.
[41] N. Wilson, N. Nghiem, H. Eyles et al., “Modeling health gains
and cost savings for ten dietary salt reduction targets,” Nu-
trition Journal, vol. 15, no. 1, p. 44, 2016.
[42] D. Klaus, J. Hoyer, and M. Middeke, “Salt restriction for the
prevention of cardiovascular disease,” Deutsches Aerzteblatt
Online, vol. 107, no. 26, pp. 457–462, 2010.
[43] J. Tuomilehto, P. Jousilahti, D. Rastenvte et al., “Urinary
sodium excretion and cardiovascular mortality in Finland: a
prospective study,” The Lancet, vol. 357, no. 9259, pp. 848–
851, 2001.
[44] WHO, “Reducing salt intake,” 21-10-2011,” 2016, http://www.
euro.who.int/en/health-topics/disease-prevention/nutrition/
news/news/2011/10/reducing-salt-intake.
[45] H. Mason, A. Shoaibi, R. Ghandour et al., “A cost effectiveness
analysis of salt reduction policies to reduce coronary heart
disease in four eastern mediterranean countries,” PLoS One,
vol. 9, no. 1, Article ID e84445, 2014.
Hindawi
https://doi.org/10.1155/2019/4059549
Research Article
The Fractional Differential Model of HIV-1 Infection of
CD4+ T-Cells with Description of the Effect of Antiviral
Drug Treatment
Bijan Hasani Lichae,1,2 Jafar Biazar ,3 and Zainab Ayati 4
1
Department of Mathematics, Guilan Science and Research Branch, Islamic Azad University, Rasht, Iran
2
Department of Mathematics, Rasht Branch, Islamic Azad University, Rasht, Iran
3
Department of Applied Mathematics, Faculty of Mathematical Sciences, University of Guilan, P.O. Box 41335-1914,
Rasht, Guilan, Iran
4
Department of Engineering Sciences, Faculty of Technology and Engineering East of Guilan, University of Guilan,
P.C. 44891-63157, Rudsar-Vajargah, Iran
Correspondence should be addressed to Jafar Biazar; biazar@guilan.ac.ir
Received 13 June 2018; Revised 11 September 2018; Accepted 17 December 2018; Published 8 January 2019
Guest Editor: Katarzyna Kolasa
Copyright © 2019 Bijan Hasani Lichae et al. This is an open access article distributed under the Creative Commons Attribution
properly cited.
In this paper, the fractional-order differential model of HIV-1 infection of CD4+ T-cells with the effect of drug therapy has been
introduced. There are three components: uninfected CD4+ T-cells, x, infected CD4+ T-cells, y, and density of virions in plasma, z.
The aim is to gain numerical solution of this fractional-order HIV-1 model by Laplace Adomian decomposition method (LADM).
The solution of the proposed model has been achieved in a series form. Moreover, to illustrate the ability and efficiency of the
proposed approach, the solution will be compared with the solutions of some other numerical methods. The Caputo sense has
been used for fractional derivatives.
1. Introduction control and prevent the spread of HIV-related diseases

[3–13]. Usually almost all of these mathematical models
Human immunodeficiency virus (HIV) is a retrovirus explain the relation between HIV viruses and uninfected
that causes acquired immunodeficiency syndrome (AIDS) CD4+ cells and the effect of drug therapy to infected cells.
[1]. HIV infects, damages, and reduces CD4+ T-cells. Bonhoeffer et al. [4] presented a model for virus dynamics
Therefore, it causes to decrease the resistance of immune with two components x and y, where x denotes the
system [2]. The body becomes more gradually sensitive to density of infected cells and y shows the density of virus-
infections and loses its safety. AIDS is one the most producing cells.
important and dangerous diseases in our time. According The proposed model is as follows:
to UNAIDS 2017 annual report, “36.7 million people dx
globally were living with HIV and 1.8 million people ⎪
⎧
⎪ � c − βx − cxy,
⎪
⎪
became newly infected with HIV and 1 million people ⎨ dt
⎪ (1)
died from AIDS-related illnesses in 2016.” In spite of the ⎪
⎪
great progress in controlling the disease, no vaccine has ⎩ dy � cxy − dy,
⎪
dt
been yet discovered for HIV. In the last two decades, a lot
of efforts have been made to design and solve mathe- where c is the rate of production of infected cells, β
matical models that have essential rule in analyzing to is the natural death rate of infected cells, d is the rate of
virus-producing cells’ death, and c is the rate of infection of method (GEM). In [31], equation (3) has also been solved
uninfected cells. This model and many such models were by homotopy perturbation method (HPM). Adomian
inspired from Anderson’s model [14, 15]. Anderson’s model [32], introduced a decomposition method (ADM) which
is one of the first and the most important models of in- is a powerful method to get analytic approximate solution
fectious diseases. Tuckwell and Wan [16] introduced a of differential equations. Using Laplace transform
modified model of equation (1) with three components: method with couple of ADM (LADM) to solve systems of
uninfected, infected CD4+ T-cells, and density of virions in differential equations leads to an effective method that
plasma (x, y, and z, respectively). The presented model with finds many applications in applied mathematics. In this
three equations is as follows: paper, we will solve equation (3) by LADM and will
⎪
⎧ dx compare the results with the results achieved by gener-
⎪
⎪ � s′ − μx − βxz, alized Euler, homotopy analysis, homotopy perturbation,
⎪
⎪ dt
⎪
⎪ and Runge–Kutta methods. The structure of the paper is
⎪
⎪
⎪
⎪ as follows: in Section 2, we will present a brief review of
⎪
⎨ dy
� βxz − εy, (2) fractional calculus. In Sections 3 and 4, we will solve the
⎪
⎪ dt
⎪
⎪ fractional-order HIV-1 model by LADM. In Section 5, the
⎪
⎪
⎪
⎪ convergence of the method will be discussed. In the last
⎪
⎪
⎪
⎪ dz section, we present the conclusion.
⎩ � cy − cz,
dt
with the initial conditions x(0) � k1 ,y(0) � k2 , and z(0) �
k3 where s′ , μ, β, ε, c, and c are constant coefficients, s′ is
2. Fractional Calculus
the rate of creation or production of CD4+ T-cells, μ is the The purpose of this section is to recall a few preliminaries
natural death rate, β is the rate of infected CD4+ cells about what appears in this research.
from uninfected CD4+ cells, ε is the rate of virus-
producing cells’ death, c is the rate of creation of vi-
rions viruses by infected cells, and c is the rate of virus Definition 2.1. The Riemann–Liouville fractional integral of
particle death. For the sake of comparison and showing order α for a function X : (0, ∞) ⟶ R is defined as
the ability of the proposed approach, we use the pa- 1 s
rameter values reported in references [6, 16]. The pa- Jα X(s) � 􏽚 (s − t)α−1 X(t)dt, (4)
Γ(α) 0
rameter values are as follows: s′ � 0.272 (day/mm3),
μ � 0.00136 (day/mm3), β � 0.00027 (day/virion/mm3), where αε(0, ∞) (see [33]).
ε � 0.33 (day/mm3), c � 50 (virion/CLM/day), and c � 2.0
(day). The rate of some coefficients will change if drug Definition 2.2. The Caputo fractional derivative for a
therapy is not 100% successful. When the drug treatment function X : (0, ∞) ⟶ R on the closed interval [0, S] is
begins, infected cells which create virus components are defined as
affected. If the drug therapy is not effective, a part of
1 s
infected cells will improve and remaining cells will begin Dα X(s) � 􏽚 (s − t)m−α−1 Xm (t)dt,
Ø
m � ºα + 1,
to produce virus [8]. Γ(m − α) 0
Mathematical modeling of many problems in biology (5)
and other branches of sciences appears as differential
equations in fractional order. Because the fractional- where α is the integer part of α. Another presentation
order differential equations save memory on them- of the Caputo fractional derivative can be shown as follows
selves and are related to fractals [8, 17–19], we prefer (see [33]):
to use the fractional-order form of the model (2) as Dα X(s) � Jm−α Dm X(s)􏼁. (6)
follows:
⎪
⎧
⎪ Dα1 (x) � s′ − μx − βxz,
⎪
⎨ Lemma 2.1. If αε(0, ∞) and m � α + 1, then the following
⎪ Dα2 (y) � βxz − εy, (3)
⎪
⎪ result holds for fractional calculus:
⎩ α3
D (z) � cy − cz, m−1
Xj (0) j
Jα 􏼂Dα X􏼃(s) � X(s) + 􏽘 s. (7)
with the same initial conditions, where 0 < αi ≤ 1, j�0
j!
i � 1, 2, 3. There are some numerical approaches for
solving these types of mathematical models. Some of
these methods are as follows: homotopy analysis, optimal
homotopy asymptotic, homotopy perturbation, Adomian
Proof. See [33, 34]. □
decomposition, and variational iteration [20–30]. In [8],
system of fractional equation (3) has been solved by Definition 2.3. The Laplace transform of Caputo fractional
homotopy analysis method (HAM) and generalized Euler derivative is defined as follows:
m−1 ∞
L􏼈Dα X(t)􏼉 � sα Y(s) − 􏽘 sα−k−1 Xj (0), xz � 􏽘 pi , (13)
j�0 (8) i�0
m − 1 < α < m, m ∈ N.
where pi is as the following equation:
3. Solution of Model (3) pi � 􏽘 xk zi−k , (14)

k�0
In this section, LADM has been implemented to solve
system of fractional equation (3) with the initial substituting equations (12)–(14) into (11) reads
conditions.
We apply Laplace transform on both sides of each
equation of equation (3): ⎪
⎧
⎪ k1
⎪
⎪ L x0 􏼁 � ,
⎪
⎪ s
⎪
⎧ L{Dα1 x} � L􏼈s′ − μx − βxz􏼉, ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎨ ⎪
⎪
⎪
⎪ k2
⎪ L􏼈Dα2 y􏼉 � L􏼈βxz − εy􏼉, (9) ⎪
⎪
⎪
L y0 􏼁 � ,
⎪
⎪ ⎪
⎪ s
⎪
⎩ ⎪
⎪
L{Dα3 z} � L􏼈cy − cz􏼉, ⎪
⎪
⎪
⎪ k3
⎪
⎪L z0 􏼁 � ,
⎪
⎪
which implies that ⎪
⎪ s
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎧
⎪ sα1 L{x} − sα1 −1 x(0) � L􏼈s′ − μx − βxz􏼉, ⎪
⎪
⎪
s′ μ β
⎪
⎪ ⎪
⎪ L x1 􏼁 � − L x0 􏼁 − α L p0 􏼁,
⎨ α ⎪
⎪ sα1 +1 sα1 s1
s 2 L􏼈y􏼉 − sα2 −1 y(0) � L􏼈βxz − εy􏼉, (10) ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎩ α3 ⎪
⎪ β ε
s L{z} − sα3 −1 z(0) � L􏼈cy − cz􏼉. ⎪
⎪ L y1 􏼁 � L p0 􏼁 − α L y0 􏼁,
⎪
⎪ sα 2 s2
⎪
⎪
⎪
⎪
Substitution of the initial conditions in equation (10) and ⎪
⎪ c c
⎪
⎪ L z1 􏼁 � L y0 􏼁 − α L z0 􏼁,
applying inverse Laplace transform results in ⎪
⎪ α
⎪
⎪ s 3 s3
⎪
⎪
1 ⎨
⎪
⎧
⎪ x � k1 + L−1 􏼔 α L􏼈s′ − μx − βxz􏼉􏼕, μ β (15)
⎪
⎪ s1 ⎪
⎪
⎪
⎪ ⎪
⎪ L x2 􏼁 � − L x1 􏼁 − α L p1 􏼁,
⎪
⎪ ⎪
⎪ α
s1 s1
⎪
⎪ ⎪
⎪
⎪
⎨ ⎪
⎪
1 ⎪
⎪
⎪ y � k2 + L−1 􏼔 α L􏼈βxz − εy􏼉􏼕, (11) ⎪
⎪ β ε
⎪
⎪ s 2 ⎪
⎪ L y2 􏼁 � L p1 􏼁 − α L y1 􏼁,
⎪
⎪ ⎪
⎪ sα 2 s2
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ 1 ⎪
⎪
⎩ z � k3 + L−1 􏼔 α L􏼈cy − cz􏼉􏼕. ⎪
⎪ c c
s3 ⎪L
⎪ z2 􏼁 � L y1 􏼁 − α L z1 􏼁,
⎪
⎪ sα 3 s3
⎪
⎪
⎪
⎪
To apply ADM, let us consider x, y, and z as the fol- ⎪
⎪
⎪
⎪ ⋮
lowing series: ⎪
⎪
⎪
⎪
⎪
⎪
∞ ⎪
⎪ μ β
⎪
⎪ L xn+1 􏼁 � − L xn 􏼁 − α L pn 􏼁,
x � 􏽘 xi , ⎪
⎪ sα 1 s1
⎪
⎪
i�0 ⎪
⎪
⎪
⎪
⎪
⎪ β ε
∞ ⎪
⎪ L yn+1 􏼁 � L pn 􏼁 − α L yn 􏼁,
y � 􏽘 yi , (12) ⎪
⎪ sα 2 s2
⎪
⎪
i�0 ⎪
⎪
⎪
⎪
⎪
⎪ c c
∞ ⎩L zn+2 􏼁 � L yn 􏼁 − α L zn 􏼁.
s α3 s3
z � 􏽘 zi .
i�0
To decompose the nonlinear term xz, let us follow an We take inverse Laplace transform on both sides of each
alternate algorithm [35] to get, equation of equation (15):
x0 � k1 ,
y0 � k2 ,
z0 � k3 ,
tα1
x1 � s′ − μx0 − βx0 z0 􏼁 ,
Γ α1 + 1􏼁
tα2
y1 � βx0 z0 − εy0 􏼁 ,
Γ α2 + 1􏼁
tα3
z1 � cy0 − cz0 􏼁 ,
Γ α3 + 1􏼁
t2α2 tα1 +α3

x2 � 􏼈− μ + βz0 􏼁 s′ − μx0 − βx0 z0 􏼁􏼉 − βx0 cy0 − cz0 􏼁 ,
Γ 2α2 + 1􏼁 Γ α1 + α3 + 1􏼁
tα2 +α3 tα1 +α2 t2α2

y2 � βx0 cy0 − cz0 􏼁 + βz0 s′ − μx0 − βx0 z0 􏼁 − ε βx0 z0 − εy0 􏼁 ,
Γ α2 + α3 + 1􏼁 Γ α1 + α2 + 1􏼁 Γ 2α2 + 1􏼁
tα2 +α3 t2α3

z2 � c βx0 z0 − εy0 􏼁 − c cy0 − cz0 􏼁 ,
Γ α2 + α3 + 1􏼁 Γ 2α3 + 1􏼁
tα1 +α2 +α3 tα1 +2α3 t2α1 +α2

x3 � βx0 c βx0 z0 − εy0 􏼁 + βx0 c cy0 − cz0 􏼁 + β2 x0 z0 cy0 − cz0 􏼁
Γ α1 + α2 + α3 + 1􏼁 Γ α1 + 2α3 + 1􏼁 Γ 2α1 + α2 + 1􏼁
t3α1 Γ α1 + α3 + 1􏼁
+ s′ − μx0 − βx0 z0 􏼁􏼈βz0 μ + z0 􏼁 + μ μ + βz0 􏼁􏼉 + β􏼨 cy0 − cz0 􏼁􏼠μx0 − s′ − μx0 − βx0 z0 􏼁 􏼡􏼩
Γ 3α1 + 1􏼁 Γ α1 + 1􏼁Γ α3 + 1􏼁
t2α1 +α3
· ,
Γ 2α1 + α3 + 1􏼁
t2α2 +α3 tα2 +2α3

y3 � βx0 c βx0 z0 − εy0 􏼁 − ε cy0 − cz0 􏼁􏼁􏼁 − βx0 c cy0 − cz0 􏼁
Γ 2α2 + α3 + 1􏼁 Γ α2 + 2α3 + 1􏼁
t2α1 +α2 Γ α1 + α3 + 1􏼁 tα1 +α2 +α3

− βz0 μ + βz0 􏼁 s′ − μx0 − βx0 z0 􏼁 + β s′ − μx0 − βx0 z0 􏼁
Γ 2α1 + α2 + 1􏼁 Γ α1 + 1􏼁Γ α3 + 1􏼁 Γ α1 + α2 + α3 + 1􏼁
tα1 +2α2 t3α2

− 􏼐β2 x0 z0 cy0 − cz0 􏼁 + βz0 ε s′ − μx0 − βx0 z0 􏼁􏼑 + ε2 βx0 z0 − εy0 􏼁 ,
Γ α1 + 2α2 + 1􏼁 Γ 3α2 + 1􏼁
tα2 +2α3 t3α3

z3 � βx0 c cy0 − cz0 􏼁 − cc βx0 z0 − εy0 􏼁􏼁 + c2 cy0 − cz0 􏼁
Γ α2 + 2α3 + 1􏼁 Γ 3α3 + 1􏼁
tα1 +α2 +α3 t2α2 +α3

+ βz0 c s′ − μx0 − βx0 z0 􏼁 − εc βx0 z0 − εy0 􏼁 .
Γ α1 + α2 + α3 + 1􏼁 Γ 2α2 + α3 + 1􏼁
(16)
We have calculated four terms of the infinite series of x, ⎪

⎧
∞
y, and z as an approximate solution. To get any desired ⎪
⎪ x(t) � 􏽘 xi (t) ≈ x0 (t) + x1 (t) + x2 (t) + x3 (t),
⎪
⎪
accuracy, one is able to proceed the process and obtain more ⎪
⎪ i�0
⎪
⎪ ∞
terms. Finally, the solution of mathematical model can be ⎨
⎪ y(t) � 􏽘 yi (t) ≈ y0 (t) + y1 (t) + y2 (t) + y3 (t), (17)
obtained as follows: ⎪
⎪ i�0
⎪
⎪
⎪
⎪ ∞
⎪
⎪
⎩ z(t) � 􏽘 zi (t) ≈ z0 (t) + z1 (t) + z2 (t) + z3 (t).
i�0
4. Numerical Simulation ε � 0.33 (day/mm3), c � 50 (virion/CLM/day), and c � 2.0

(day) and the initial conditions x(0) � 100, y(0) � 0, and
In this section, constants and initial values are substituted in z(0) � 1 in equation (16), we get,
equation (16) to obtain an approximate solution.
Substituting the following values: s′ � 0.272 (day/mm3),
μ � 0.00136 (day/mm3), β � 0.00027 (day/virion/mm3),
x0 � 100,
y0 � 0,
z0 � 1,
tα1
x1 � 0.1090 ,
Γ α1 + 1􏼁
tα2
y1 � 0.0270 ,
Γ α2 + 1􏼁
tα3
z1 � −2.0000 ,
Γ α3 + 1􏼁
t2α2 tα1 +α3

x2 � −0.0001776700 + 0.05400 ,
Γ 2α2 + 1􏼁 Γ α1 + α3 + 1􏼁
tα2 +α3 tα1 +α2 t2α2

y2 � −0.05400 + 0.0000294300 − 0.0089100 ,
Γ α2 + α3 + 1􏼁 Γ α1 + α2 + 1􏼁 Γ 2α2 + 1􏼁
tα2 +α3 t2α3

z2 � 1.35000 + 4.0000 ,
Γ α2 + α3 + 1􏼁 Γ 2α3 + 1􏼁
tα1 +α2 +α3 tα1 +2α3 t2α1 +2α2

x3 � −0.0364500 − 0.10800 − 0.0000145800
Γ α1 + α2 + α3 + 1􏼁 Γ α1 + 2α3 + 1􏼁 Γ 2α1 + α2 + 1􏼁
t3α1 Γ α1 + α3 + 1􏼁 t2α1 +α3

+ 0.000029711656 − 0.00054 ,
Γ 3α1 + 1􏼁 Γ α1 + 1􏼁Γ α3 + 1􏼁 Γ 2α1 + α3 + 1􏼁
t2α2 +α3 tα2 +2α3 t2α1 +α2

y3 � 0.054270000 + 0.10800 − 0.000000047971
Γ 2α2 + α3 + 1􏼁 Γ α2 + 2α3 + 1􏼁 Γ 2α1 + α2 + 1􏼁
Γ α1 + α3 + 1􏼁 tα1 +α2 +α3 tα1 +2α3 t3α2

+ 0.0000294300 + 0.000004868100 + 0.0002940300 ,
Γ α1 + 1􏼁Γ α3 + 1􏼁 Γ α1 + α2 + α3 + 1􏼁 Γ α1 + 2α3 + 1􏼁 Γ 3α2 + 1􏼁
tα2 +2α3 t3α3 tα1 +α2 +α3

z3 � 0.054270000 − 8.0000 + 0.0014715000
Γ α2 + 2α3 + 1􏼁 Γ 3α3 + 1􏼁 Γ α1 + α2 + α3 + 1􏼁
t2α2 +α3
− 0.4455000 .
Γ 2α2 + α3 + 1􏼁
(18)
Three terms approximations can be written as the fol-

lowing form:
t α1 t2α2 tα1 +α3 tα1 +α2 +α3

x(t) � 100 + 0.1090 − 0.0001776700 + 0.05400 − 0.0364500
Γ α1 + 1􏼁 Γ 2α2 + 1􏼁 Γ α 1 + α3 + 1 􏼁 Γ α1 + α2 + α3 + 1􏼁
tα1 +2α3 t2α1 +α2 t3α1

− 0.10800 − 0.0000145800 + 0.000029711656
Γ α1 + 2α3 + 1􏼁 Γ 2α1 + α2 + 1􏼁 Γ 3α1 + 1􏼁
Γ α1 + α3 + 1􏼁 t2α1 +α3
− 0.00054 ,
Γ α1 + 1􏼁Γ α3 + 1􏼁 Γ 2α1 + α3 + 1􏼁
t α2 tα2 +α3 tα1 +α2 t2α2

y(t) � 0.0270 − 0.05400 + 0.0000294300 − 0.0089100
Γ α2 + 1􏼁 Γ α2 + α3 + 1􏼁 Γ α1 + α2 + 1􏼁 Γ 2α2 + 1􏼁
t2α2 +α3 tα2 +2α3 t2α1 +α2

+ 0.054270000 + 0.10800 − 0.000000047971 (19)
Γ 2α2 + α3 + 1􏼁 Γ α2 + 2α3 + 1􏼁 Γ 2α1 + α2 + 1􏼁
Γ α1 + α3 + 1􏼁 tα1 +α2 +α3 tα1 +2α2

+ 0.0000294300 + 0.000004868100
Γ α1 + 1􏼁Γ α3 + 1􏼁 Γ α1 + α2 + α3 + 1􏼁 Γ α1 + 2α2 + 1􏼁
t3α2
+ 0.0002940300 ,
Γ 3α2 + 1􏼁
t α3 tα2 +α3 t2α3 tα2 +2α3

z(t) � 1 − 2.0000 + 1.35000 + 4.0000 + 0.054270000
Γ α3 + 1􏼁 Γ α2 + α3 + 1􏼁 Γ 2α3 + 1􏼁 Γ α2 + 2α3 + 1􏼁
t3α3 tα1 +α2 +α3 t2α2 +α3

− 8.0000 + 0.0014715000 − 0.4455000 .
Γ 3α3 + 1􏼁 Γ α1 + α2 + α3 + 1􏼁 Γ 2α2 + α3 + 1􏼁
Let us take α1 , α2 , and α3 equal to α, so the solution of

fractional-order of model (3) is obtained as follows:
⎪
⎧
⎪ 0.1090tα 0.0538223300t2α 0.1444348683t3α 0.00054Γ(2α + 1)t3α
⎪
⎪ x(t) � 100 + + − − ,
⎪
⎪ Γ(α + 1) Γ(2α + 1) Γ(3α + 1) Γ(3α + 1)
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨ 0.0270tα 0.0628805700t2α 0.1625688501t3α 0.0000294300Γ(2α + 1)t3α
⎪ y(t) � − + + , (20)
⎪
⎪ Γ(α + 1) Γ(2α + 1) Γ(3α + 1) Γ(3α + 1)
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ 2.0000tα 5.35000t2α 8.389758500t3α
⎩ z(t) � 1 − + − .
Γ(α + 1) Γ(2α + 1) Γ(3α + 1)
For α1 � α2 � α3 � 1, the solution of equation (3) will be In Tables 1–3, one can compare the approximate solu-
as follows: tion of fractional-order of model (3) with the results of GEM,
x(t) � 100 + 0.1090t + 0.02691116500t2 − 0.02425247806t3 , HAM, RK4 in [8], and HPM in [31] using traditional order
⎪
⎧
⎪
⎨ α � 1. The results of LADM are more accurate than the
⎪ y(t) � 0.0270t − 0.03144028500t2 + 0.02710461836t3 , results obtained by other methods.
⎪
⎩
z(t) � 1 − 2.0000t + 2.6750000t2 − 1.398293083t3 . Figures 1–3 show the results for different values of α, and
(21) the results can be compared.
Table 1: Numerical results of x(t) (uninfected CD4+ T-cells).

t LADM GEM HPM HAM RK4
0 100 100 100 100 100
0.2 100.023 100.023 100.023 100.023 100.023
0.4 100.046 100.047 100.047 100.047 100.047
0.6 100.070 100.071 100.071 100.071 100.071
0.8 100.092 100.097 100.097 100.096 100.097
1.0 100.112 100.122 100.123 100.122 100.122
Table 2: Numerical results of y(t) (infected CD4+ T-cells).

0 0 0 0 0 0
0.2 0.00436 0.00434 0.00434 0.00434 0.00434
0.4 0.00750 0.00715 0.00721 0.00714 0.00715
0.6 0.01074 0.00908 0.00934 0.00909 0.00908
0.8 0.01336 0.01049 0.01117 0.01063 0.01049
1.0 0.01866 0.01161 0.01276 0.01194 0.01161
Table 3: Numerical results of z(t) (density of virions in plasma).

0 1 1 1 1 1
0.2 0.69581 0.69030 0.69071 0.69059 0.69070
0.4 0.53851 0.51152 0.51208 0.51237 0.51190
0.6 0.46097 0.41069 0.41394 0.40994 0.41103
0.8 0.39607 0.35656 0.37749 0.35148 0.35684
1.0 0.27671 0.33053 0.42419 0.32869 0.33073
100.10
100.08
x(t) values
100.06
100.04
100.02
100
0 0.2 0.4 0.6 0.8 1
t values
α=1 α = 0.5
α = 0.75 α = 0.25
Figure 1: Dynamics of uninfected CD4+ T-cells for various values of α.
5. Convergence Analysis of the Method where F is a functional operator and can be decomposed as
F � Dα + R + N and g is a known function.
In this section, convergence of the proposed method, using Dα is a Caputo fractional derivative operator, R is a linear
the idea presented in [36], is studied. operator, and N is a nonlinear analytic operator, re-
Consider the following functional equation: spectively. So equation (22) can be written as follows:
F(v(t)) � g(t), (22) Dα (v(t)) � g(t) − R(v(t)) − N(v(t)). (23)
0.12
0.10
0.08
x(t) values
0.06
0.04
0.02
0
0 0.2 0.4 0.6 0.8 1
t values
α=1 α = 0.5
α = 0.75 α = 0.25
Figure 2: Dynamics of infected CD4+ T-cells for various values of α.
0.8
0.6
x(t) values
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
t values
α=1 α = 0.5
α = 0.75 α = 0.25
Figure 3: Dynamics of density of virions in plasma for various values of α.
The goal is to find a function v(t) satisfying equation which implies that
(22). By applying the Laplace transform on both sides of L􏼈g(t)􏼉 −1 L{R(v(t))}
equation (23) reads v(t) � v0 + L−1 􏼨 α
􏼩−L 􏼨 􏼩
s sα
L􏼈Dα (v(t))􏼉 � L􏼈g(t) − R(v(t)) − N(v(t))􏼉. (24) (27)
L{N(v(t))}
−1
By using definition (2.3), equation (23) can be written as −L 􏼨 􏼩.
sα
follows:
v(0) L􏼈g(t)􏼉 L{R(v(t))} L{N(v(t))} By implementing ADM and assuming the solution v(t)
L{v(t)} � + − − , as an infinite series say, v(t) � 􏽐∞
n�0 vn (t), and writing the
s sα sα sα nonlinear term based on Adomian polynomials such as
(25)
∞
by considering v(0) � v0 and using inverse of Laplace N(v(t)) � 􏽘 pn v0 (t), v1 (t), . . . , vn (t)􏼁, (28)
transform on both sides of equation (25) results in n�0
v(0) L􏼈g(t)􏼉 where

L−1 {L{v(t)}} � L−1 􏼨 +
s sα 1 dn ⎧ ⎨ n ⎬
⎫
(26) pn v0 (t), v1 (t), . . . , vn (t)􏼁 � ⎝
⎛ ⎠
i⎞
n ⎩ Nλ 􏽘 vi (t)λ ⎭ .
L{R(v(t))} L{N(v(t))} Γ(n + 1) dλ i�0 λ�0
− − 􏼩,
sα sα (29)
Equation (27) can be written as the following form: ∞

L􏼈g(t)􏼉 ∞
L􏼈R vn (t)􏼁􏼉
􏽘 vn (t) � v0 + L−1 􏼨 􏼩 − 􏽘 L−1 􏼨 􏼩
∞
L􏼈g(t)􏼉 L 􏼈R 􏽐∞
n�0 vn (t)􏼁􏼉 s α sα
􏽘 vn (t) � v0 + L−1 􏼨 α
􏼩 − L−1 􏼨 􏼩 n�0 n�0
n�0 s sα ∞
L􏼈pn v0 (t), v1 (t), . . . , vn (t)􏼁􏼉
− 􏽘 L−1 􏼨 􏼩.
L􏼈􏽐∞
n�0 pn v0 (t), v1 (t), . . . , vn (t)􏼁􏼉 sα
− L−1 􏼨 􏼩. n�0
sα (31)
(30)
From which we can define
So, we have
⎧
⎪ v0 (t) � v0 ,
⎪
⎪
⎪
⎪
⎪
⎪ L􏼈g(t)􏼉 L􏼈R v0 (t)􏼁􏼉 L􏼈p0 v0 (t)􏼁􏼉
⎪
⎪ v1 (t) � L−1 􏼨 􏼩 − L−1 􏼨 􏼩 − L−1 􏼨 􏼩,
⎪
⎪ s α s α sα
⎪
⎪
⎪
⎪
⎪
⎨ L􏼈R v1 (t)􏼁􏼉 L􏼈p1 v0 (t), v1 (t)􏼁􏼉
⎪ v2 (t) � −L−1 􏼨 􏼩 − L−1 􏼨 􏼩, (32)
⎪
⎪ s α sα
⎪
⎪
⎪
⎪
⎪
⎪ ⋮
⎪
⎪
⎪
⎪
⎪
⎪ L􏼈R vn (t)􏼁􏼉 L􏼈pn v0 (t), v1 (t), . . . , vn (t)􏼁􏼉
⎪
⎩ vn+1 (t) � −L−1 􏼨 􏼩 − L−1 􏼨 􏼩.
s α sα
Theorem 5.1. LADM for equation (21) with the solution (19) Then, by assumption of equation (33), we have
is equivalent to L􏼈g(t)􏼉 −1 L􏼈R v0 (t)􏼁􏼉
sn (t) � v0 (t) + v1 (t) + . . . + vn (t), v1 (t) � L−1 􏼨 􏼩−L 􏼨 􏼩
(33) sα sα
s0 (t) � v0 (t). (37)
By using the following iterative scheme: L􏼈p0 v0 (t)􏼁􏼉
−1
−L 􏼨 􏼩,
L􏼈g(t)􏼉 sα
−1 L􏼈R sn (t)􏼁􏼉
sn+1 (t) � v0 + L−1 􏼨 α
􏼩−L 􏼨 􏼩
s sα for n � 1:
L􏼈N sn (t)􏼁􏼉 L􏼈g(t)􏼉 −1 L􏼈R s1 (t)􏼁􏼉
− L−1 􏼨 􏼩, s2 (t) � v0 + L−1 􏼨 α
􏼩−L 􏼨 􏼩
sα s sα
(34) L􏼈N s1 (t)􏼁􏼉
− L−1 􏼨 􏼩
where sα
n n
⎝􏽘 v (t)⎞
N⎛ ⎠ � 􏽘 p v (t), v (t), . . . , v (t)􏼁, L􏼈g(t)􏼉 −1 L􏼈R v0 (t) + v1 (t)􏼁􏼉
i i 0 1 i
(35) � v0 + L−1 􏼨 α
􏼩−L 􏼨 􏼩
i�0 i�0 s sα
n � 0, 1, 2, . . . .
L􏼈p0 v0 (t)􏼁 + p1 v0 (t), v1 (t)􏼁􏼉
− L−1 􏼨 􏼩
sα
Proof. For n � 0, from equation (33):
L􏼈R v1 (t)􏼁􏼉
L􏼈g(t)􏼉
−1 −1 L􏼈R s0 (t)􏼁􏼉 � v0 (t) + v1 (t) − L−1 􏼨 􏼩
s1 (t) � v0 + L 􏼨 􏼩−L 􏼨 􏼩 sα
s α sα
L􏼈p1 v0 (t), v1 (t)􏼁􏼉
− L−1 􏼨 􏼩.
L􏼈N s0 (t)􏼁􏼉
−1 sα
−L 􏼨 􏼩
sα (38)
L􏼈g(t)􏼉 −1 L􏼈R v0 (t)􏼁􏼉 We know that s2 (t) � v0 (t) + v1 (t) + v2 (t), so we obtain
� v0 + L−1 􏼨 􏼩−L 􏼨 􏼩
sα sα L􏼈R v1 (t)􏼁􏼉 −1 L􏼈p1 v0 (t), v1 (t)􏼁􏼉
v2 (t) � −L−1 􏼨 􏼩−L 􏼨 􏼩.
sα sα
L􏼈p0 v0 (t)􏼁􏼉
− L−1 􏼨 􏼩. (39)
sα
(36) By strong induction, let us have
L􏼈R vk (t)􏼁􏼉 Proof.

vk+1 (t) � −L−1 􏼨 􏼩 ��
sα ��sn+1 − sn �� vn+1 �� ≤ c��vn �� ≤ c2 ��vn−1 �� ≤ . . . ≤ cn+1 ��v0 ��.
(44)
L􏼈pk v0 (t), v1 (t), . . . , vk (t)􏼁􏼉 (40)
− L−1 􏼨 􏼩,
sα ∀n, m ∈ N, n ≥ m, we have
��
��sn − sm �� sn − sn−1 􏼁 + sn−1 − sn−2 􏼁 + . . . + sm+1 − sm 􏼁��
k � 1, 2, . . . , n − 1, ��
≤ ��sn − sn−1 �� + ��sn−1 − sn−2 �� + . . . + ��sm+1 − sm ��
and prove the following for k � n, ��
≤ cn ��v0 �� + cn−1 ��v0 �� + . . . + cm+1 ��v0 ��
L􏼈g(t)􏼉 −1 L􏼈R sn (t)􏼁􏼉
sn+1 (t) � v0 + L−1 􏼨 􏼩−L 􏼨 􏼩 ��
s α sα ≤ 􏼐cn + cn−1 + . . . + cm+1 􏼑��v0 ��
L􏼈N sn (t)􏼁􏼉 cm+1 ��

− L−1 􏼨 􏼩 ≤ cm+1 􏼐1 + c + c2 + . . . + cn + . . .􏼑 ≤ �v �.
sα 1−c 0
n
(45)
L􏼈g(t)􏼉 −1 L􏼈R 􏽐k�0 vk (t)􏼁􏼉
� v0 + L−1 􏼨 􏼩 − L 􏼨 􏼩 This means that lim ‖sn − sm ‖ � 0, therefore, 􏼈sn 􏼉 is a
sα sα n,m ⟶∞
Cauchy sequence in the Banach space of X and is conver-
L􏼈􏽐nk�0 pk v0 (t), v1 (t), . . . , vk (t)􏼁􏼉 gent. So, ∃ s ϵ X, s.t lim sn � s.
− L−1 􏼨 􏼩 n ⟶∞
sα From equation (33), we derive
n L􏼈g(t)􏼉
L􏼈g(t)􏼉 L􏼈R vk (t)􏼁􏼉 lim sn (t) � v0 + L−1 􏼨 􏼩
� v0 + L−1 􏼨 􏼩 − 􏽘 L−1 􏼨 􏼩 n ⟶∞ sα
sα k�0
sα
⎪
⎧ lim R sn (t)􏼁􏼛⎫
⎪
n ⎨ L􏼚n ⟶∞
⎪ ⎪
⎬
L􏼈pk v0 (t), v1 (t), . . . , vk (t)􏼁􏼉
−1 − L−1 ⎪ ⎪
− 􏽘L 􏼨 􏼩 ⎪
⎩ s α ⎪
⎭
k�0
sα
⎪
⎧ lim N sn (t)􏼁􏼛⎫
⎪
L􏼈R vn (t)􏼁􏼉 ⎨ L􏼚n ⟶∞
⎪ ⎪
⎬
� v0 (t) + v1 (t) + . . . vn (t) − L−1 􏼨 􏼩 − L−1 ⎪ ⎪
sα ⎪
⎩ s α ⎪
⎭
L􏼈pn v0 (t), v1 (t), . . . , vn (t)􏼁􏼉 L􏼈g(t)􏼉
− L−1 􏼨 􏼩. � v0 + L−1 􏼨 􏼩
sα sα
(41) ⎪
⎧ ⎪
⎪ lim R 􏽐nk�0 vk (t)􏼁􏼛⎫
⎨ L􏼚n ⟶∞ ⎪
⎬
−1
Then, from equation (33), we derive −L ⎪ α ⎪
⎪
⎩ s ⎪
⎭
L􏼈R vn (t)􏼁􏼉
vn+1 (t) � −L−1 􏼨 􏼩 ⎪
⎧ ⎪
sα ⎪ lim N 􏽐nk�0 vk (t)􏼁􏼛⎫
⎨ L􏼚n ⟶∞ ⎪
⎬
(42) − L−1 ⎪ α ⎪
⎪
⎩ s ⎪
⎭
L􏼈pn v0 (t), v1 (t), . . . , vn (t)􏼁􏼉
−1
−L 􏼨 􏼩.
sα L􏼈g(t)􏼉
� v0 + L−1 􏼨 􏼩
sα
This entails the statement is true and the theorem is
proved. □ ⎧
⎪ lim 􏽐nk�0 R vk (t)􏼁􏼛⎫
⎨ L􏼚n ⟶∞
⎪ ⎪
⎪
⎬
−1
−L ⎪ α ⎪
⎪
⎩ s ⎪
⎭
Theorem 5.2. Let X be a Banach space.
⎪
⎧ lim 􏽐nk�0 pk v0 (t), v1 (t), . . . , vk (t)􏼁􏼛⎫
⎪
(i) 􏽐∞i�0 vi (t) resulted from equation (31), convergence to
⎨ L􏼚n ⟶∞
⎪ ⎪
⎬
− L−1 ⎪ ⎪
s ∈ X, if ∃ c ∈ [0, 1), s.t (∀ n ∈ N ⟹ ‖vn+1 ‖ ≤ c ‖vn ‖), ⎪
⎩ s α ⎪
⎭
(ii) s(t) � 􏽐∞ i�0 vi (t) satisfies in
L􏼈g(t)􏼉
� v0 + L−1 􏼨 􏼩
sα
L􏼈g(t)􏼉 −1 L{R(s(t))} L􏼈􏽐∞
k�0 R vk (t)􏼁􏼉
s(t) � v0 + L−1 􏼨 􏼩−L 􏼨 􏼩 − L−1 􏼨 􏼩
s α sα sα
(43) L􏼈􏽐∞
k�0 pk v0 (t), v1 (t), . . . , vk (t)􏼁􏼉
L{N(s(t))} − L−1 􏼨 􏼩,
− L−1 􏼨 sα
􏼩.
sα (46)
From equation (35), we have then Dα x(t) ⟶ Dx(t); therefore, the fractional-order of
∞ ∞ presented model reduces to traditional model. By applying
⎝􏽘 v (t)⎞
N⎛ ⎠ � 􏽘 p v (t), v (t), . . . , v (t)􏼁, Laplace transform and Adomian decomposition method, or
i i 0 1 i
i�0 i�0 (47) LADM for short, which is a strong approach to compute
numerical solution of fractional differential equations, we
n � 0, 1, 2. . . . .
gain an approximate solution of the proposed model. The
So, accuracy of the proposed approach has made it a reliable
∞ method. We have calculated four terms of the infinite series
L􏼈g(t)􏼉 −1 L􏼈R 􏽐k�0 vk (t)􏼁􏼉 of x, y, and z as an approximate solution. The result of
s(t) � v0 + L−1 􏼨 􏼩 − L 􏼨 􏼩
sα sα LADM has been compared with the results of some other
methods such as GEM, HAM, RK4 [8], and HPM [31]. The
L􏼈N 􏽐∞
k�0 vk (t)􏼁􏼉 results are presented in Tables 1–3. Figures 1–3, show that
− L−1 􏼨 􏼩
sα the uninfected CD4+ T-cells, x, infected CD4+ T-cells, y, and
density of virions in plasma, z, depend on the various values
L􏼈g(t)􏼉 −1 L{R(s(t))} of α, the various values of the parameters, and the time
� v0 + L−1 􏼨 􏼩−L 􏼨 􏼩 fractional derivative. A comparison of the approximate
sα sα
solutions shows that LADM can work more accurate than
L{N(s(t))} other methods. Convergence of the proposed method is
− L−1 􏼨 􏼩. studied. Because of the fact that obtaining the exact solution
sα
for system of fractional equation is difficult or impossible, we
(48) would like to suggest such an easy and reliable approach for
□ further research, in the future.
Lemma 5.1. Equation (43) is equivalent to equation (22). Data Availability

Proof. By using Laplace transform on both sides of equation Data used to support this study are available at DOI:
(43) reads to https://doi.org/10.1038/sj.icb.7100056. These prior studies
L􏼈g(t)􏼉 L{R(s(t))} L{N(s(t))} (and datasets) are cited at relevant places within the text as
L(s(t)) � L v0 􏼁 + − − references [6, 8, 16].
sα sα sα
v(0) L􏼈g(t)􏼉 L{R(s(t))} L{N(s(t))} Disclosure
� + − −
s sα sα sα
This research did not receive specific funding, but has
sα−1 v(0) L􏼈g(t)􏼉 L{R(s(t))} L{N(s(t))} resulted from a Ph.D. program, by Mr. Bijan Hasani Lichaee,
� + − − ,
sα sα sα sα under supervision of Prof. Jafar Biazar and cosupervision of
(49) Dr. Zainab Ayati.
so we can write
Conflicts of Interest
sα L(s(t)) − sα−1 v(0) � L􏼈g(t)􏼉 − L{R(s(t))}
(50) The authors declare that they have no conflicts of interest.
− L{N(s(t))}.
In virtue of definition 2.3 and linearity of the Laplace References

transform, equation (50) can be written as the follows:
L􏼈Dα s(t)􏼉 � L􏼈g(t) − R(s(t)) − N(s(t))􏼉. (51) [1] R. Weiss, “How does HIV cause AIDS?,” Science, vol. 260,
no. 5112, pp. 1273–1279, 1993.
By applying the inverse of Laplace transform on both [2] D. C. Douek, M. Roederer, and R. A. Koup, “Emerging
sides of equation (51), we derive concepts in the immunopathogenesis of AIDS,” Annual Re-
view of Medicine, vol. 60, no. 1, pp. 471–484, 2009.
Dα s(t) � g(t) − R(s(t)) − N(s(t)). (52) [3] A. V. Herz, S. Bonhoeffer, R. M. Anderson, R. M. May, and
M. A. Nowak, “Viral dynamics in vivo: limitations on esti-
Considering v(t) � s(t), one gets equation (22). So, the mates of intracellular delay and virus decay,” Proceedings of
solution of equation (43) is the same as the solution of the National Academy of Sciences, vol. 93, no. 14, pp. 7247–
equation (22). □ 7251, 1996.
[4] S. Bonhoeffer, J. M. Coffin, and M. A. Nowak, “Human
immunodeficiency virus drug therapy and virus load,” Journal
6. Conclusion of Virology, vol. 71, no. 4, pp. 3275–3278, 1997.
[5] R. V. Culshaw and S. Ruan, “A delay-differential equation
In this paper, a fractional-order model of HIV-1 with model of HIV infection of CD4+ T-cells,” Mathematical
three components has been introduced. When α ⟶ 1, biosciences, vol. 165, no. 1, pp. 27–39, 2000.
[6] H. C. Tuckwell and F. Y. M. Wan, “On the behavior of so- [24] M. Merdan, A. Gökdoğan, and A. Yildirim, “On the numerical
lutions in viral dynamical models,” BioSystems, vol. 73, no. 3, solution of the model for HIV infection of CD4+ T cells,”
pp. 157–161, 2004. Computers and Mathematics with Applications, vol. 62, no. 1,
[7] C. Lv and Z. Yuan, “Stability analysis of delay differential pp. 118–123, 2011.
equation models of HIV-1 therapy for fighting a virus with [25] Y. Khan, H. Vazquez-Leal, and Q. Wu, “An efficient iterated
another virus,” Journal of Mathematical Analysis and Ap- method for mathematical biology model,” Neural Computing
plications, vol. 352, no. 2, pp. 672–683, 2009. and Applications, vol. 23, no. 3-4, pp. 677–682, 2013.
[8] A. A. M. Arafa, S. Z. Rida, and M. Khalil, “The effect of anti- [26] Ş. Yüzbaşı, “A numerical approach to solve the model for HIV
viral drug treatment of human immunodeficiency virus infection of CD4+ T cells,” Applied Mathematical Modelling,
type 1 (HIV-1) described by a fractional order model,” vol. 36, no. 12, pp. 5876–5890, 2012.
Applied Mathematical Modelling, vol. 37, no. 4, pp. 2189– [27] M. Ghoreishi, A. M. Ismail, and A. Alomari, “Application of
2196, 2013. the homotopy analysis method for solving a model for HIV
[9] A. A. M. Arafa, S. Z. Rida, and M. Khalil, “Fractional infection of CD4+ T-cells,” Mathematical and Computer
modeling dynamics of HIV and CD4+ T-cells during pri- Modelling, vol. 54, no. 11-12, pp. 3007–3015, 2011.
mary infection,” Nonlinear biomedical physics, vol. 6, no. 1, [28] A. Atangana and E. F. D. Goufo, “On the mathematical
p. 1, 2012. analysis of Ebola hemorrhagic fever: deathly infection disease
[10] Y. Yan and C. Kou, “Stability analysis for a fractional dif- in West African countries,” BioMed Research International,
vol. 2014, Article ID 261383, 7 pages, 2014.
ferential model of HIV infection of CD4+ T-cells with time
[29] M. Y. Ongun, “The Laplace adomian decomposition method
delay,” Mathematics and Computers in Simulation, vol. 82,
for solving a model for HIV infection of CD4+ T cells,”
no. 9, pp. 1572–1585, 2012.
Mathematical and Computer Modelling, vol. 53, no. 5-6,
[11] L. Wang and M. Y. Li, “Mathematical analysis of the global
pp. 597–603, 2011.
dynamics of a model for HIV infection of CD4+ T cells,”
[30] A. A. M. Arafa, S. Rida, and M. Khalil, “A fractional-order
Mathematical Biosciences, vol. 200, no. 1, pp. 44–57, 2006. model of HIV infection with drug therapy effect,” Journal of
[12] A. S. Perelson, D. E. Kirschner, and R. De Boer, “Dynamics of the Egyptian Mathematical Society, vol. 22, no. 3, pp. 538–543,
HIV infection of CD4+ T cells,” Mathematical Biosciences, 2014.
vol. 114, no. 1, pp. 81–125, 1993. [31] M. Merdan and T. Khan, “Homotopy perturbation method
[13] J. Biazar, “Solution of the epidemic model by Adomian de- for solving viral dynamical model,” CÜ Fen-Edebiyat
composition method,” Applied Mathematics and Computa- Fakültesi, Fen Bilimleri Dergisi, vol. 31, pp. 65–77, 2010.
tion, vol. 173, no. 2, pp. 1101–1106, 2006. [32] G. Adomian, Nonlinear Stochastic Systems Theory and Ap-
[14] R. M. Anderson and R. M. May, “Population biology of in- plications to Physics, Springer Science & Business Media, Vol.
fectious diseases: part I,” Nature, vol. 280, no. 5721, 46, Springer Science & Business Media, Berlin, Germany,
pp. 361–367, 1979. 1988.
[15] R. M. May and R. M. Anderson, “Population biology of in- [33] K. Diethelm, The Analysis of Fractional Differential Equations:
fectious diseases: part II,” Nature, vol. 280, no. 5722, An Application-Oriented Exposition using Differential Oper-
pp. 455–461, 1979. ators of Caputo Type, Springer, Berlin, Germany, 2010.
[16] H. C. Tuckwell and F. Y. M. Wan, “Nature of equilibria and [34] A. A. Kilbas, H. M. Srivastava, and J. J. Trujillo, Preface,
effects of drug treatments in some simple viral population Elsevier, New York City, NY, USA, 2006.
dynamical models,” IMA Journal of Mathematical Control [35] J. Biazar, E. Babolian, G. Kember, A. Nouri, and R. Islam, “An
and Information, vol. 17, no. 4, pp. 311–327, 2000. alternate algorithm for computing Adomian polynomials in
[17] N. Heymans and J.-C. Bauwens, “Fractal rheological models special cases,” Applied Mathematics and Computation,
and fractional differential equations for viscoelastic behavior,” vol. 138, no. 2-3, pp. 523–529, 2003.
Rheologica Acta, vol. 33, no. 3, pp. 210–219, 1994. [36] Z. Ayati and J. Biazar, “On the convergence of Homotopy
[18] M. Giona and H. E. Roman, “Fractional diffusion equation on perturbation method,” Journal of the Egyptian Mathematical
fractals: one-dimensional case and asymptotic behaviour,” Society, vol. 23, no. 2, pp. 424–428, 2015.
Journal of Physics A: Mathematical and General, vol. 25, no. 8,
pp. 2093–2105, 1999.
[19] F. B. Tatom, “The relationship between fractional calculus and
fractals,” Fractals, vol. 3, no. 1, pp. 217–229, 1995.
[20] S. J. Liao, The proposed homotopy analysis technique for the
solution of nonlinear problems, Ph.D. thesis, Shanghai Jiao
Tong University, Shanghai, China, 1992.
[21] V. S. Ertürk, Z. M. Odibat, and S. Momani, “An approximate
solution of a fractional order differential equation model of
human T-cell lymphotropic virus I (HTLV-I) infection of
CD4+ T-cells,” Computers and Mathematics with Applica-
tions, vol. 62, no. 3, pp. 996–1002, 2011.
[22] Y. Ding and H. Ye, “A fractional-order differential equation
model of HIV infection of CD4+ T-cells,” Mathematical and
Computer Modelling, vol. 50, no. 3-4, pp. 386–392, 2009.
[23] F. Haq, K. Shah, G. U. Rahman et al., “Numerical solution of
fractional order smoking model via laplace Adomian de-
composition method,” Alexandria Engineering Journal,
vol. 57, no. 2, pp. 1061–1069, 2017.
Hindawi
https://doi.org/10.1155/2018/4091497
Research Article
Control of Blood Glucose for Type-1 Diabetes by Using
Reinforcement Learning with Feedforward Algorithm
Phuong D. Ngo ,1 Susan Wei,2 Anna Holubová,3 Jan Muzik,3 and Fred Godtliebsen1
1
UiT The Arctic University of Norway, Tromsø, Norway
2
The University of Melbourne, Australia
3
Czech Technical University, Prague, Czech Republic
Correspondence should be addressed to Phuong D. Ngo; phuong.ngo@uit.no
Received 9 August 2018; Revised 18 November 2018; Accepted 28 November 2018; Published 30 December 2018
Copyright © 2018 Phuong D. Ngo et al. This is an open access article distributed under the Creative Commons Attribution License,
Background. Type-1 diabetes is a condition caused by the lack of insulin hormone, which leads to an excessive increase in blood
glucose level. The glucose kinetics process is difficult to control due to its complex and nonlinear nature and with state
variables that are difficult to measure. Methods. This paper proposes a method for automatically calculating the basal and
bolus insulin doses for patients with type-1 diabetes using reinforcement learning with feedforward controller. The algorithm
is designed to keep the blood glucose stable and directly compensate for the external events such as food intake. Its per-
formance was assessed using simulation on a blood glucose model. The usage of the Kalman filter with the controller was
demonstrated to estimate unmeasurable state variables. Results. Comparison simulations between the proposed controller
with the optimal reinforcement learning and the proportional-integral-derivative controller show that the proposed
methodology has the best performance in regulating the fluctuation of the blood glucose. The proposed controller also
improved the blood glucose responses and prevented hypoglycemia condition. Simulation of the control system in different
uncertain conditions provided insights on how the inaccuracies of carbohydrate counting and meal-time reporting affect the
performance of the control system. Conclusion. The proposed controller is an effective tool for reducing postmeal blood
glucose rise and for countering the effects of external known events such as meal intake and maintaining blood glucose at a
healthy level under uncertainties.
1. Introduction By using control theories, various studies have been

conducted to design a control system for patients with type-1
Type-1 diabetes is a chronic condition that is characterized diabetes. For example, Marchetti et al. [2], derived an im-
by an excessive increase in blood glucose level because the proved proportional-integral-derivative controller for blood
pancreas does not produce insulin hormone due to the glucose control. Soylu et al. [3] proposed a Mamdani type
autoimmune destruction of pancreatic beta cells. High blood fuzzy control strategy for exogenous insulin infusion. Model
glucose can lead to both acute and chronic complications predictive control has also been widely used in type-1 di-
and eventually result in failure of various organs. abetes and artificial pancreas development [4, 5]. Recently,
Until today, there are many challenges in control of the together with the development of artificial intelligence and
blood glucose in type-1 diabetes. One of them is that the machine learning, reinforcement learning (RL) has emerged
glucose kinetics process is complex, nonlinear, and only as a data-driven method to control unknown nonlinear
approximately known [1]. There are also many external systems [6, 7] and has been used as a long-term management
known and unknown factors that affect the blood glucose tool for chronic diseases [8, 9]. The biggest advantage of RL
level such as food intakes, physical activities, stress, and compared to other methods is that the algorithm depends
hormone changes. Generally, it is difficult to predict and only on interactions with the system and does not require a
quantify those factors and disturbances. well represented model of the environment. This especially
makes RL well suited for type-1 diabetes since the modelling estimating unmeasurable state variables during the control
process of the insulin-kinetic dynamics is complex and process. This paper also conducts simulations under un-
requires invasive measurements on the patient or must be fit certain information to evaluate the robustness of the pro-
through a large dataset. Hence, by using RL as the control posed controller.
algorithm, the modelling process can be bypassed, which
makes the algorithm not susceptible to any modelling error.
In diabetes, controlling of blood glucose require actions
2. Methods
that are made at specific instance throughout the day in 2.1. Problem Formulation. The purpose of our study is to
terms of insulin doses or food intakes. The actions are based design an algorithm to control the blood glucose in patients
on the current observable states of the patients (e.g., blood with type-1 diabetes by the means of changing the insulin
glucose measurement and heart rate). The effectiveness of concentration. The blood glucose metabolism is a dynamic
the actions is calculated by how far the measured blood system in which the blood glucose changing over time as the
glucose value is compared to the healthy level. In RL, an results of many factors such as food intake, insulin doses,
agent makes decision based on the current state of the physical activities, and stress level. The learning process of
environment. The task of the algorithm is to maximize a RL is based on the interaction between a decision-making
cumulative reward function or to minimize a cumulative agent and its environment, which will lead to an optimal
cost function. Based on these similarities in the decision- action policy that results in desirable states [17]. The RL
making process between a human being and a RL agent, RL framework for type-1 diabetes includes the following
may be key to the development of an artificial pancreas elements:
system.
When dealing with meal disturbances, modelling of (i) The state vector at time instance k consists of the
glucose ingestion is the norm as well as the first step in states of the patient:
designing a controller for disturbance rejection [10]. Feed-
forward control was proven to be an effective tool to improve
disturbance rejection performance [11, 12]. In control sys- xk � 􏼂 g(k) − gd (k) χ(k) 􏼃T , (1)
tem theory, feed-forward is the term that describes a con-
where g(k) and gd (k) the are measured and desired blood
troller that utilizes the signal obtained when there is a (large)
glucose levels, respectively, and χ(k) is the interstitial insulin
deviation from the model. Compared to feed-back control,
activity (defined in the appendix).
where action is only taken after the output has moved away
from the setpoint, the feed-forward architecture is more (ii) The control variable (insulin action) uk , which is part
proactive since it uses the disturbance model to suggest the of the total insulin ik (a combination of the basal and
time and size of control action. Furthermore, building a meal the bolus insulin (Figure 1)):
disturbance model is simpler and requires less data to fit
than finding the insulin-glucose kinetics. Based on the
model, necessary changes in insulin actions can be calculated ik � ubasal (k) + ubolus (k) � uk + ubasal (0) + ubolus (k), (2)
to compensate for the effects of carbohydrate on the blood
glucose level. where ubasal (k) and ubolus (k) are the basal and bolus at time
A challenge in the control of the blood glucose is the lack instance k, respectively.
of real-time measurement techniques. With the develop- (iii) The cost received one time-step later as a conse-
ment of continuous glucose measurement sensors, blood quence of the action. In this paper, the cost was
glucose level can be measured and provided to the controller calculated by the following quadratic function:
in minute intervals. However, blood glucose value alone is
usually not enough to describe the states of the system for
control purpose. Therefore, an observer is needed to estimate rk+1 � xTk Qxk + uTk Ruk , (3)
other variables in the state space from the blood glucose
measurement. In this paper, the Kalman filter was chosen for 1 0
where Q � 􏼢 􏼣 and R � 0.01. Each element in matrix
that purpose since it can provide an optimal estimation of 0 0.1
the state variables when the system is subjected to process Q and the value of R indicate the weighting factors of the
and measurement noises [13, 14]. cost function. The element in the first row and the first
Vrabie et al. [15] established methodologies to obtain column of Q has the highest value, which corresponds to the
optimal adaptive control algorithms for dynamical systems weighting of the difference between the measured blood
with unknown mathematical models by using reinforcement glucose and the prescribed healthy value. Since our ultimate
learning. Based on that, Ngo et al. [16] proposed a re- goal is to reduce this difference, the factor of this mea-
inforcement learning algorithm for updating basal rates in surement should have the largest value in the cost function.
patients with type-1 diabetes. This paper completes the The element in the second row and second column of Q
framework for blood glucose control with both basal and corresponds to the weighting of the interstitial insulin ac-
bolus insulin doses. The framework includes the re- tivity. The value of R indicates the weighting factor of the
inforcement learning algorithm, the feed-forward controller action (basal update). Minimizing the cost function,
for compensating food intake and the Kalman filter for therefore, becomes the problem of minimizing the difference
The action-value function depends on the current state

D D D2 D2 and the next action. It was shown that the action-value
Carb G function satisfies the following recursive equation (Bellman
Gmeal chi
equation) [15, 17]:
D u_bolus Gglucose
Qπk (x, u) � rk + cQπk+1 (x, u). (6)
ubolus
Gff +
+ I chi Since the state space and action space are infinite,
ubasal
function approximation was used in this paper for esti-
Gins mation of the Q-function. In this case, the Q-function was
x xhat G
u_basal approximated as a quadratic function of vectors xk and uk :
Q Kalman filter
x Qπk (x, u) ≈ zTk Pzk , (7)
Actor Q
u
Critic
where the symmetric and positive definite matrix P is called
the kernel matrix and contains the parameters that need to
be estimated. Vector zk is the combined vector of xk and uk :
Figure 1: Control system diagram. T
z � 􏽨 xTk uTk 􏽩 . (8)
between the measured blood glucose and the desired value, With Kronecker operation, the approximated Q-func-
the interstitial insulin activity, and the change in basal tion can be expressed as a linear combination of the basis
insulin. function Φ(zk ) � zk ⊗ zk :
At time instance k + 1, a sequence of observations would Qπk (x, u) ≈ zTk Pzk � wT zk ⊗ zk 􏼁 � wT Φ zk 􏼁, (9)
be xk , uk , rk+1 , xk+1 and uk+1 . Based on this observation, the
agent receives information about the state of the patient and where w is the vector that contains elements of P and ⊗ is the
chooses an insulin action. The body reacts to this action and Kronecker product.
transitions to a new state. This determines the cost of the By substituting Qπk (x, u) in equation (6) by wT Φ(zk ) and
action. using the policy iteration method with the least square al-
For the control design purpose, the blood glucose model gorithm [15], elements of vector w can be estimated. Matrix
(Appendix) was divided into three submodels: the meal P can then be obtained from w using the Kronecker
(Gmeal), the insulin (Gins), and the glucose kinetics (Gglucose). transformation.
The controller has three main components: the actor, the By decomposing the kernel matrix P into smaller ma-
critic, and the feedforward algorithm. The actor is used to trices Pxx , Pxu , Pux , and Puu , the approximated Q-function
estimate the action-value function, the critic’s task is to can be written as follows:
obtain optimal basal insulin, and the feedforward algorithm T T
1⎡⎢⎢ xk ⎤⎥⎥ ⎡⎢⎢ xk ⎤⎥⎥ 1⎡⎢⎢ xk ⎤⎥⎥ ⎡⎢⎢ Pxx Pxu ⎥⎥⎤⎡ xk
is used to propose the bolus insulin profile for disturbance Qπk (x, u) � ⎣ ⎦ P ⎣ ⎦ � ⎣ ⎦ ⎣ ⎦⎢
⎣ ⎤⎥⎥⎦.
⎢
compensation (food intake). The purpose of the Kalman 2 u uk 2 u Pux Puu uk
k k
filter is to estimate unmeasurable states of the patient.
(10)
The current policy is improved with actions that min-
2.2. Basal Update by Actor and Critic. When the patient is in imize the Q-function Qπk (x, u). This can be done by first
a fasting condition, the controller only needs to change the taking the partial derivative of the Q-function and then
basal insulin level through the actor and the critic. Based on solving zQπk (x, u)/zu � 0. The optimal solution can there-
the current state xk , the actor proposes an insulin action uk after be obtained as follows [15]:
through the policy π : uk � π(xk ). The updated basal rate is
obtained from uk as follows: uk � −P−1
uu Pux xk . (11)
ubasal � uk + ubasal (0). (4) With that, the update of basal insulin is
After each action, the patient transforms into a new state, ubasal � −P−1
uu Pux xk + ibe , (12)
and the cost associated with the previous action can be
where ibe is the equilibrium basal plasma insulin
calculated using equation (3). The action-value function (Q-
concentration.
function) of action u is defined as the accumulation of cost
when the controller takes action uk � u at time instance k
and then continues following policy π(xk+1 ):
2.3. Bolus Update by Feedforward Algorithm. When the
⎨∞
⎧ ⎬
⎫ patient consumes meals, in addition to the basal insulin, the
Qπk (x, u) � Eπ ⎩ 􏽘 ci rk+i+1 ∣ xk � x, uk � u⎭ , (5) controller calculates and applies boluses to compensate for
i�0
the rise of blood glucose as the results of carbohydrate in the
where c (with 0 < c ≤ 1) is the discount factor that indicates food. The feedforward algorithm first predicts how much
the weighting of future cost in the action-value function. blood glucose level will rise and then suggests a bolus profile
to counter the effects of the meal. The starting time of the Table 1: Parameters and constants of the insulin-glucose kinetics
bolus doses was also calculated by the algorithm based on the model.
meal intake model. Name Description Value
Since the meal intake model (equations (A.1) and (A.2))
p1 Glucose effectiveness 0.2 min−1
and the insulin model (equation (A.4)) are linear time- p2 Insulin sensitivity 0.028 min−1
invariant (LTI) models, they can be transformed from p3 Insulin rate of clearance 10−4 min−1
state space equations into transfer functions as follows: AG Carbohydrate bioavailability 0.8 min−1
D (s) −1 AG τD Glucose absorption constant 10 min
Gmeal (s) � 2 � Cmeal sI − Ameal 􏼁 Bmeal � 2, V Plasma volume 2730 g
D(s) sτ D + 1􏼁 Equilibrium basal plasma insulin 7.326 μIU/
ibe
concentration ml
D (s) −1 p3
Gins (s) � 2 � Cins sI − Ains 􏼁 Bins � ,
D(s) s + p2
(13) Table 2: Variables of the insulin-glucose kinetics model.
Name Description Unit
where
D Amount of CHO intake mmol/min
−1/τ D 0 D1 Amount of glucose in compartment 1 mmol
Ameal � 􏼢 􏼣,
1/τ D −1/τ D D2 Amount of glucose in compartment 2 mmol
g(t) Plasma glucose concentration mmol/l
AG χ(t) Interstitial insulin activity min−1
Bmeal � 􏼢 􏼣, i(t) Plasma insulin concentration μIU/ml
0
(14)
Cmeal � 􏼂 0 1/τ D 􏼃,
Ains � −p2 , compartments 1 and 2 cannot be measured directly during
implementation, Kalman filter was used to provide an es-
Bins � p3 , timation of the state variables from the blood glucose level.
Cins � 1. The discretized version of the type-1 diabetes system can be
written in the following form:
Descriptions and values of τ D , p2 , and p3 are shown in
Tables 1 and 2. The transfer function from the meal intake xK (k + 1) � AK xK (k) + BK uK (k) + HK w(k),
D(s) to the blood glucose level g(s) can be calculated as (19)
yK (k) � CK xK (k) + v(k),
g(s)
F(s) � � Gmeal (s) + Gff (s)Gins (s)􏼁Gglucose (s). T
where xK (k) � 􏼂 D1 D2 g(k) − gd (k) χ(k) 􏼃 , uK (k) �
D(s) T
(15) 􏼂 D(k) i(k) 􏼃 , and matrices AK , BK , CK are linearized
coefficient matrices of the model:
In order to compensate for the meal, the gain of the open
loop system F(s) must be made as small as possible. Hence, −1/τ D 0 0 0
⎡⎢⎢⎢ ⎤⎥⎥
the feedforward transfer function was chosen such that ⎢⎢⎢ 1/τ −1/τ
⎢ D D 0 0 ⎥⎥⎥⎥⎥
Gmeal (s) + Gff (s)Gins (s) ⟶ 0, which leads to AK � ⎢⎢⎢⎢ ⎥⎥⎥,
⎢⎢⎢ 0
⎣ 1/τ D −p1 − gd 0 ⎥⎥⎥⎥⎦
−AG s + p2 􏼁
Gff (s) � −Gmeal (s)G−1
ins (s) � 2. (16) 0 0 0 −p2
p3 τ D s + 1􏼁
AG 0 (20)
The meal compensation bolus in s-domain can be cal- ⎡⎢⎢⎢ ⎤⎥⎥
⎢⎢⎢ 0
⎢ 0 ⎥⎥⎥⎥⎥
culated from the feedforward transfer function: BK � ⎢⎢⎢⎢ ⎥⎥⎥,
⎢⎢⎢ 0 0 ⎥⎥⎥
−AG s + p2 􏼁 ⎣ ⎥⎦
ubolus (s) � Gff (s)D(s) � 2 D(s). (17)
p3 τ D s + 1􏼁 0 p3 V
CK � 􏼂 0 0 1 0 􏼃,
Hence, the feedforward action becomes the output of the
T
following dynamic system, which can be solved easily using matrix HK is the noise input matrix: HK � 􏼂 0 0 0 p3 V 􏼃 ,
any ordinary differential equation solver: the output value yK (k) � g(k) − gd (k) is the measured
p3 τ 2D u€bolus (t) + 2p3 τ D u_ bolus (t) + p3 ubolus (t) blood glucose deviation from the desired level, w(k) is the
(18) insulin input noise, and v(k) is the blood glucose mea-
_
� −AG 􏼐D(t) + p2 D(t)􏼑. surement noise with zero-mean Gaussian distribution. The
variances of w(k) and v(k) are assumed to be as follows:
E􏼐w2 (k)􏼑 � Rw ,
2.4. Kalman Filter for Type-1 Diabetes System. Since the (21)
interstitial insulin activity, the amounts of glucose in E􏼐v2 (k)􏼑 � Rv .
Based on the discretized model, a Kalman filter was rates in the simulations are characterized by parameter τ D
implemented through the following equation: from the model, where τ D � 50 corresponds to food with a
x􏽢(k + 1 ∣ k) � Ak · x􏽢(k ∣ k − 1) + Bk · uK (k) slow absorption rate and τ D � 10 corresponds to food with a
(22) fast absorption rate. The amount of carbohydrate (CHO) per
+ L[y(k) − C · x􏽢(k ∣ k − 1)], meal can be found in Figure 2.
Next, the performance of the proposed controller was
where x􏽢(k + 1 ∣ k) denotes the estimation of x(k + 1) based evaluated under uncertainties of meal information. Two
on measurements available at time k. The gain L is the cases of uncertainties were considered: uncertain CHO es-
steady-state Kalman filter gain, which can be calculated by timation case and uncertain meal-recording time. In the
−1
L � MCT 􏼐CMCT + Rv 􏼑 , (23) uncertain CHO estimation, the estimated CHO information
that provided to the controller was assumed to be a normal
where M is the solution of the corresponding algebraic distribution with a standard deviation of 46% from the
Riccati equation [13, 14, 18]: correct carbohydrate value shown in Figure 2. The standard
−1 deviation value was used based on the average adult esti-
M � AMAT + BRw B − AMCT 􏼐CMCT + R􏼑 CMAT . mates and the computerized evaluations by the dietitian
(24) [19]. For the uncertain meal-recording time, the estimated
meal starting time is assumed to be a normal distribution
By assuming the noise variances to be Rw � Rv � 0.01, with a standard deviation of two minutes from the real
the Kalman filter gain was calculated from equation (23) as starting time. This standard deviation value was randomly
L � 􏼂 0 0 8.32 · 10−4 −6.40 · 10−7 􏼃. (25) selected because systematic research on the accuracy of
meal-time recording for patients with type-1 diabetes could
not be found. For each case, multiple simulations were
conducted in the same closed-loop system with its corre-
2.5. Simulation Setup. First, a pretraining of the algorithm sponding random variables. From the obtained results, the
was conducted on the type-1 diabetes model in the scenario mean and standard deviation for blood glucose responses at
where the patient is in a fasting condition (without food each time point will be calculated and analyzed.
intake). The purpose of the pretraining simulation is to
obtain an initial estimation of the action-value function for
the algorithm. The learning process was conducted by re- 3. Results
peating the experiment multiple times (episodes). Each After pretraining in the no-meal scenario, the Q-function
episode starts with an initial blood glucose of 90 mg/dL and was estimated as follows:
ends after 30 minutes. The objective of the algorithm is to
search and explore actions that can drive the blood glucose 4.454 · 102 −8.870 · 104 −0.084
to its target level of 80 mg/dL. ⎡⎢⎢⎢ ⎤⎥⎥⎥
Qπk (x, u) � 􏽨 xTk uTk 􏽩⎢⎢⎢⎢⎣ −8.870 · 104 3.538 · 107 33.630 ⎥⎥⎥⎥⎦
By using the initial estimation of the action-value
function, the controller was then tested in the daily sce- −0.084 33.630 0.010
nario with food intake. Comparisons were made between the xk
proposed reinforcement learning with the feedforward ·􏼢 􏼣.
(RLFF) controller, the optimal RL (ORL) controller [15], and uk
the proportional-integral-derivative (PID) controller. The (28)
ORL was designed with the same parameters and pretrained
The initial basal policy was obtained from the initial Q-
in the same scenario as with the RLFF. The PID controller
function and equation (12):
gains were chosen, which produces a similar blood glucose
settling time as the RLFF: ubasal (k) � 8.86(g(k) − 80) − 3534.11χ(k) + 7.326. (29)
uk � Kp g(k) − gd (k)􏼁 + Ki 􏽘 g(k) − gd (k)􏼁 The initial estimation of the Q-function and the initial
k (26) basal policy were used for subsequent testing simulations of
+ Kd (g(k) − g(k − 1)), the control algorithm.
During the simulation with correct meal information,
where blood glucose responses of the RLFF, the ORL, and the PID
Kp � 1, are shown in Figures 3 and 4. The insulin concentration
during the process can also be found in Figures 5 and 6. With
Ki � 0.001, (27) slow-absorption food, the fluctuation range of blood glucose
Kd � 0.01. was approximately ±30 mg/dL for all three controllers from
the desired value (Figure 3). However, with fast absorption
In order to understand the effects of different food types glucose meals, the fluctuation range of the postmeal blood
on the controlled system, two sets of simulations were glucose level was within ±40 mg/dL with the RLFF compared
conducted for food that has slow and fast glucose absorption to ±60 mg/dL with the ORL and is significantly smaller than
rates while containing a similar amount of carbs. Absorption the fluctuation range ±80 mg/dL of the PID (Figure 4).
60 25
Insulin action, uk (mU/L)

50 20
CHO level (g)
40
15
30
20 10
10 5
0 0
8 10 12 14 16 18 20 22 24 0 5 10 15 20
Time (hour) Time (hour)
Figure 2: CHO consumed throughout the day. RLFF
ORL
150 PID
Blood glucose level (mg/dL)
Figure 5: Comparison of insulin concentrations in the nominal

condition for slow glucose absorption food.
100 120
Insulin action, uk (mU/L)

100
80
50 60
0 5 10 15 20 40
Time (hour)
20
RLFF
0
ORL
PID 0 5 10 15 20
Time (hour)
Figure 3: Comparison of the blood glucose responses in the
nominal condition for slow glucose absorption food. RLFF
ORL
PID
180 Figure 6: Comparison of insulin concentrations in the nominal
160 condition for fast glucose absorption food.

140
120 fast glucose absorption food and 5 mg/dL for slow glucose
100 absorption food.
80
60
4. Discussion
40
0 5 10 15 20 The controller has shown its capability to reduce the rise of
Time (hour) postmeal blood glucose in our simulations. It can be seen in
RLFF
Figures 3 and 4 that three controllers were able to stabilize
ORL the blood glucose. However, when using the RLFF, the added
PID bolus makes the insulin responses much faster when there is
a change in blood glucose level, which reduces the peak of
Figure 4: Comparison of the blood glucose responses in the
the postmeal glucose rise by approximately 30 percent
nominal condition for fast glucose absorption food.
compared to the ORL and 50 percent compared to the PID
in the fast-absorption case. It can also be seen that the
Figures 7 and 8 show the blood glucose variation under undershoot blood glucose (the distance between the lowest
uncertain meal time and CHO counting. The upper and blood glucose and the desired blood glucose value) of the
lower bounds in shaded areas show the mean blood glucose PID controller is much larger than that of the RLFF and the
value plus and minus the standard deviation for each in- ORL. The RLFF has the smallest glucose undershoot among
stance. Under uncertain meal information, the upper the three controllers. Low blood glucose value (hypogly-
bound was kept to be smaller than 40 mg/dL from the cemia) can be very dangerous for patients with type-1 di-
desired blood glucose value for fast glucose absorption food abetes. Therefore, simulation results show the advantage of
and 15 mg/dL for slow glucose absorption food. The lower using RLFF in improving safety for patients. In general, with
bound is smaller than 15 mg/dL from the desired value for the feedforward algorithm, the proposed algorithm is an
130
120

110
100
90
80
70
8 10 12 14 16 18 20 22 24
Time (hour)
Uncertain CHO counting

Uncertain meal time
Figure 7: Blood glucose responses under uncertainties for fast glucose absorption food.
130 1 diabetes. The controller regulates the patient’s glucose level

120 using both basal and bolus insulin. Simulation results of the
proposed controller, the optimal reinforcement learning,
110
and the PID controller on a type-1 diabetes model show that
100 the proposed algorithm is the most effective algorithm. The
90 basal updates can stabilize the blood glucose, and the bolus
80
can reduce the glucose undershoot and prevent hypogly-
cemia. Comparison of the blood glucose variation under
70
different uncertainties provides understandings of how the
8 10 12 14 16 18 20 22 24 accuracy of carbohydrate estimation and meal-recording
Time (hour) time can affect the closed-loop responses. The results
Uncertain CHO counting show that the control algorithm was able to keep the blood
Uncertain meal time glucose at a healthy level although uncertainties create
variations in the blood glucose responses.
Figure 8: Blood glucose responses under uncertainties for slow
glucose absorption food.
Appendix
effective tool for countering the effects of external events Blood Glucose Model
such as meal intake.
Among uncertainties, carb counting created more effect In this paper, the insulin-glucose process was used as the
on the variation of the blood glucose than meal-time re- subject in our simulations. The model is described by the
cording, especially with slow absorbing food. The un- following equations [20–23]:
certainty in recording meal time may also lead to larger dD1 (t) D (t)
undershot of blood glucose below the desired level as can be � AG D(t) − 1 , (A.1)
dt τD
seen in Figure 7. Following the same trend as previous
simulations, the fluctuation range of the blood glucose with dD2 (t) D1 (t) D2 (t)
slow absorbing food is smaller than the fluctuation range � − , (A.2)
dt τD τD
with fast glucose absorbing food. In general, the control
algorithm kept the blood glucose at the healthy level al-
dg(t) D (t)
though uncertainties affect the variation of the responses. � −p1 g(t) − χ(t)g(t) + 2 , (A.3)
However, an accurate carbohydrate counting and accurate dt τD
meal-time recording method are still important for the
purpose of blood glucose control in order to completely dχ(t)
� −p2 χ(t) + p3 V i(t) − ube 􏼁, (A.4)
avoid the chance of getting hypoglycemia. dt
where variable descriptions and parameter values are given
5. Conclusion in Tables 1 and 2. In this model, the inputs are the amount of
CHO intake D and the insulin concentration i(t). The output
The paper proposes a blood glucose controller based on of the model is the blood glucose concentration g(t). It is
reinforcement learning and feedforward algorithm for type- assumed that the blood glucose is controlled by using an
insulin pump, and there is no delay between the adminis- [9] J. Pineau, M. G. Bellemare, A. J. Rush, A. Ghizaru, and
tered insulin and the plasma insulin concentration. S. A. Murphy, “Constructing evidence-based treatment
strategies using methods from computer science,” Drug and
Alcohol Dependence, vol. 88, no. S2, pp. S52–S60, 2007.
Abbreviations [10] K. Lunze, T. Singh, M. Walter, M. D. Brendel, and
S. Leonhardt, “Blood glucose control algorithms for type 1
RL: Reinforcement learning
diabetic patients: a methodological review,” Biomedical Signal
RLFF: Reinforcement learning with feedforward algorithm Processing and Control, vol. 8, no. 2, pp. 107–119, 2013.
ORL: Optimal reinforcement learning [11] S. P. Bhattacharyyta, “Disturbance rejection in linear sys-
PID: Proportional-integral-derivative tems,” International Journal of Systems Science, vol. 5, no. 7,
LTI: Linear time-invariant pp. 633–637, 1974.
CHO: Carbohydrate. [12] H. Zhong, L. Pao, and R. de Callafon, “Feedforward control
for disturbance rejection: model matching and other
Data Availability methods,” in Proceedings of 24th Chinese Control and Decision
Conference (CCDC), pp. 3528–3533, Taiyuan, China, May
The data used to support the findings of this study are 2012.
available from the corresponding author upon request. [13] F. Lewis, Optimal Estimation, John Wiley & Sons, Inc.,
Hoboken, NJ, USA, 1986.
[14] G. F. Franklin, J. D. Powell, and M. L. Workman, Digital
Conflicts of Interest Control of Dynamic Systems, Addison-Wesley, Boston, MA,
USA, 2nd edition, 1990.
The authors declare no conflicts of interest. [15] D. Vrabie, K. G. Vamvoudakis, and F. L. Lewis, Optimal
Adaptive Control and Differential Games by Reinforcement
Acknowledgments Learning Principles, Institution of Engineering and Tech-
nology, vol. 81, 1st edition, 2012.
The research has been funded by financial support from [16] P. D. Ngo, S. Wei, A. Holubova, J. Muzik, and F. Godtliebsen,
Tromsø Forskningsstiftelse. The publication charges for this “Reinforcement-learning optimal control for type-1 diabetes,”
article have been funded by a grant from the publication in Proceedings of 2018 IEEE EMBS International Conference
fund of UiT, the Arctic University of Norway. on Biomedical & Health Informatics (BHI), pp. 333–336, Las
Vegas, NV, USA, March 2018.
[17] R. Sutton and A. Barto, Reinforcement Learning: An
References Introduction, MIT Press, Cambridge, MA, USA, 1st edition,
1998.
[1] Q. Wang, P. Molenaar, S. Harsh et al., “Personalized state- [18] MathWorks, MATLAB Optimization Toolbox: User’s Guide
space modeling of glucose dynamics for type 1 diabetes using (r2018a), MathWorks, Natick, MA, USA, 2018.
continuously monitored glucose, insulin dose, and meal in- [19] A. S. Brazeau, H. Mircescu, K. Desjardins et al., “Carbohydrate
take,” Journal of Diabetes Science and Technology, vol. 8, no. 2, counting accuracy and blood glucose variability in adults with
pp. 331–345, 2014. type 1 diabetes,” Diabetes Research and Clinical Practice,
[2] G. Marchetti, M. Barolo, L. Jovanovic, H. Zisser, and vol. 99, no. 1, pp. 19–23, 2013.
D. E. Seborg, “An improved PID switching control strategy [20] R. N. Bergman, Y. Z. Ider, C. R. Bowden, and C. Cobelli,
for type 1 diabetes,” IEEE Transactions on Biomedical Engi- “Quantitative estimation of insulin sensitivity,” American
neering, vol. 55, no. 3, pp. 857–865, 2008. Journal of Physiology-Endocrinology and Metabolism, vol. 236,
[3] S. Soylu, K. Danisman, I. E. Sacu, and M. Alci, “Closed-loop no. 6, p. E667, 1979.
control of blood glucose level in type-1 diabetics: a simulation [21] R. Hovorka, V. Canonico, L. J. Chassin et al., “Nonlinear
study,” in Proceedings of International Conference on Electrical model predictive control of glucose concentration in subjects
and Electronics Engineering (ELECO), pp. 371–375, Bursa, with type 1 diabetes,” Physiological Measurement, vol. 25,
Turkey, November 2013. no. 4, pp. 905–920, 2004.
[4] D. Boiroux, A. K. Duun-Henriksen, S. Schmidt et al., “Overnight [22] M. E. Wilinska, L. J. Chassin, H. C. Schaller, L. Schaupp,
glucose control in people with type 1 diabetes,” Biomedical T. R. Pieber, and R. Hovorka, “Insulin kinetics in type-1
Signal Processing and Control, vol. 39, pp. 503–512, 2018. diabetes: continuous and bolus delivery of rapid acting in-
[5] H. Lee and B. W. Bequette, “A closed-loop artificial pancreas sulin,” IEEE Transactions on Biomedical Engineering, vol. 52,
based on model predictive control: human-friendly identifi- no. 1, pp. 3–12, 2005.
cation and automatic meal disturbance rejection,” Biomedical [23] A. Mösching, Reinforcement Learning Methods for Glucose
Signal Processing and Control, vol. 4, no. 4, pp. 347–354, 2009. Regulation in Type 1 Diabetes, Ecole Polytechnique Federale
[6] M. K. Bothe, L. Dickens, K. Reichel et al., “The use of re- de Lausanne, Lausanne, Switzerland, 2016.
inforcement learning algorithms to meet the challenges of an
artificial pancreas,” Expert Review of Medical Devices, vol. 10,
no. 5, pp. 661–673, 2014.
[7] M. De Paula, L. O. Ávila, and E. C. Martı́nez, “Controlling
blood glucose variability under uncertainty using re-
inforcement learning and Gaussian processes,” Applied Soft
Computing, vol. 35, pp. 310–332, 2015.
[8] C. J. C. H. Watkins and P. Dayan, “Technical note:
Q-learning,” in Reinforcement Learning, pp. 55–68,
vol. 292Springer US, Boston, MA, USA, 1992.
Hindawi
https://doi.org/10.1155/2018/6287913
Research Article
Evaluation of the Feasibility of Screening Tau Radiotracers
Using an Amyloid Biomathematical Screening Methodology
1,2 1,2
Ying-Hwey Nai and Hiroshi Watabe
1
Division of Radiation Informatics for Medical Imaging, Graduate School of Biomedical Engineering, Tohoku University,
Sendai, Japan
2
Division of Radiation Protection and Safety Control, Cyclotron and Radioisotope Center, Tohoku University, Sendai, Japan
Correspondence should be addressed to Hiroshi Watabe; hwatabe@tohoku.ac.jp
Received 16 August 2018; Revised 6 November 2018; Accepted 21 November 2018; Published 19 December 2018
Copyright © 2018 Ying-Hwey Nai and Hiroshi Watabe. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
The purpose of this study is to evaluate the feasibility of extending a previously developed amyloid biomathematical screening
methodology to support the screening of tau radiotracers during compound development. 22 tau-related PET radiotracers were
investigated. For each radiotracer, in silico MLogP, Vx, and in vitro KD were input into the model to predict the in vivo K1, k2, and
BPND under healthy control (HC), mild cognitive impaired (MCI), and Alzheimer’s disease (AD) conditions. These kinetic
parameters were used to simulate the time activity curves (TACs) in the target regions of HC, MCI, and AD and a reference region.
Standardized uptake value ratios (SUVR) were determined from the integrated area under the TACs of the target region over the
reference region within a default time window of 90–110 min. The predicted K1, k2, and BPND values were compared with the
clinically observed values. The TACs and SUVR distributions were also simulated with population variations and noise. Finally,
the clinical usefulness index (CUI) ranking was compared with clinical comparison results. The TACs and SUVR distributions
differed for tau radiotracers with lower tau selectivity. The CUI values ranged from 0.0 to 16.2, with 6 out of 9 clinically applied tau
radiotracers having CUI values higher than the recommend CUI value of 3.0. The differences between the clinically observed
TACs and SUVR results showed that the evaluation of the clinical usefulness of tau radiotracer based on single target binding
could not fully reflect in vivo tau binding. The screening methodology requires further study to improve the accuracy of screening
tau radiotracers. However, the higher CUI rankings of clinically applied tau radiotracers with higher signal-to-noise ratio
supported the use of the screening methodology in radiotracer development by allowing comparison of candidate radiotracers
with clinically applied radiotracers based on SUVR, with respect to binding to a single target.
1. Introduction the brain occurs before clinical symptoms appear, the im-
aging of these precursors can support differential diagnosis
Alzheimer’s disease (AD) is a progressive neurodegenerative and early intervention to increase the success rate of treating
disorder defined by histopathological features such as senile AD or slow down the rate of dementia. As such, the 2018
plaques and neurofibrillary tangles (NFT), and clinical National Institute on Aging-Alzheimer’s Association (NIA-
symptoms such as memory loss and reduced executive AA) research framework includes not only symptomatic
functions [1]. The yearly number of AD cases is increasing stages of AD, but also biomarker classification involving
worldwide, leading to an increased cost of care for dementia amyloid, tau, and neurodegeneration AT(N) biomarkers [2].
patients. Positron emission tomography (PET) using amy- The new framework will be able to identify subjects at risk
loid and tau radiotracers can measure the amyloid and tau for AD for suitable and early treatment, in particular,
loads, in terms of standardized uptake values ratio (SUVR), preclinical AD subjects (classified as A+T−(N−) or
and their distributions in a subject’s brain from static PET A+T+(N+)), who are not cognitively impaired but have
images. Since abnormal accumulation of amyloid and tau in abnormal amyloid and tau protein deposits [2].
Despite active efforts since 2000 to develop amyloid and was reported to have reduced off-target binding but further
tau-targeting PET radiotracers to assist in the diagnosis of evaluation was still required [16].
AD and to support AD drug development, there are few We have previously developed an amyloid bio-
radiotracers that have made it into clinical studies and mathematical screening methodology to support the screen-
displayed good clinical efficacy. In conventional radiotracer ing of candidate amyloid radiotracers during compound
and drug development, poor bench-to-bedside translation development [4, 5]. The screening methodology predicts the
often results due to the differences between in vitro and in standardized uptake values ratios (SUVRs) of different subject
vivo conditions. Similarly, animal models, especially ro- conditions of a radiotracer and then compares the clinical
dents, are often poor predictors of human physiology and usefulness of multiple radiotracers simultaneously in dis-
treatment response and have been reported to be incorrect in criminating the subject conditions using a clinical usefulness
approximately one out of three cases [3]. Although larger index (CUI). The CUI was developed to objectively evaluate
animals (e.g., pigs and primates) show closer physiology to the clinical usefulness of a radiotracer, based on its binding
that of human, they are still in-prefect human models and capability to a single target of interest, in terms of SUVR. The
are costly for high-throughput screening compared to ro- SUVR is a semiquantitative parameter that generalizes the
dents. These issues lead to high attrition rates in drug and complicated behaviors of tau radiotracers. SUVR is also
radiotracer development. Biomathematical simulation can generally preferred for diagnosis of patients in amyloid and
complement high-throughput screening by allowing si- tau imaging; hence, the clinical data are more readily available
multaneous and rapid evaluation of many candidate ra- for comparison. Thus, we chose SUVR over other kinetic
diotracers [4–6]. parameters such as nondisplaceable binding potential (BPND,
Compared to amyloid radiotracers, the development of a unitless).
successful tau radiotracer encounters additional challenges In this study, we evaluate the feasibility of extending the
due to the tau phenotypes. Tau proteins have six isoforms, amyloid-validated screening methodology to support the
which differ in the number of exons (0, 1, 2) on the acidic development of tau PET radiotracers, where more challenges
region and the number of repeats (3 repeats (3R) or 4R) in like off-target binding exist. This is the first in silico method
the repeat-domain regions [7]. The different isoforms un- investigated, which uses the physicochemical and pharma-
dergo several posttranslational modifications, leading to cological properties of the compounds to support tau PET
various ultrastructural conformations, which will affect the radiotracers developments. 22 PET radiotracers reported to
binding of tau radiotracers. In addition, they also need to bind to tau proteins were investigated, including 9 clinically
discriminate between the paired helical filament (PHF) tau applied and tau-focused radiotracers, namely, [18F]THK523,
from other β-sheet structured aggregates such as amyloid- [18F]THK5105, [18F]THK5117, [18F]THK5317, [18F]THK5351,
beta (Aβ) and α-synuclein. Although the tau protein is larger [18F]flortaucipir, [18F]T808, [11C]PBB3 and [18F]MK6240, and
than the Aβ protein, the tau binding sites are present in 3 clinically applied but non-tau-focused radiotracers, specif-
smaller concentrations compared to the Aβ binding sites by ically [18F]Lansoprazole, [11C]Astemizole, and [18F]FDDNP.
5–20 folds; hence, the selectivity of tau over other β-sheet
structured aggregates needs to be high to ensure accurate 2. Materials and Methods
quantification. Moreover, as tau proteins exist in-
tracellularly, tau radiotracers not only need to cross the An overview of the amyloid biomathematical methodology
blood-brain barrier (BBB), they also need to be able to cross is described briefly, followed by the screening of tau PET
the cell membrane [8]. radiotracers using the biomathematical methodology. The
Existing clinically applied tau radiotracers showed some details of the methodology are found in somewhere [4, 5].
limitations. [11C]PBB3 has high binding selectivity to tau
over Aβ, but it is difficult to synthesize as it will undergo 2.1. Biomathematical Screening Methodology. The screening
photoisomerization [9]. Moreover, it is rapidly metabolized methodology was based on a simplified 1-tissue-
in the plasma, and its polar metabolite is shown to cross the compartment model (1TCM), with the assumption that
blood-brain barrier and enter into the brain [10]. The short the radiotracers cross the blood-brain barrier (BBB) by
half-life of carbon-11 has also prompted the development of passive diffusion. It consists of four main parts (Figure 1).
fluorinated PBB3 compounds ([18F]AM-PBB3 and [18F]PM-
PBB3) and other tau radiotracers so that they can be used in
hospitals without dedicated cyclotron facilities. [18F]T808 2.2. Generation of Physicochemical and Pharmacological
(also known as [18F]AV-680) exhibits defluorination, which Parameters. A total of three inputs were required for each
will affect the quantitative analysis of PET images especially radiotracer: in silico molecular volume and lipophilicity as
for regions near the skull [11]. Some THK compounds represented by McGowan Volume (Vx, cm3/mol/100),
(Tohoku University, Japan) showed differences in the uptake Moriguchi LogP (MLogP, unitless), and an in vitro disso-
due to the enantiomeric properties of the compounds [12]. A ciation constant (KD, nM) (Table 1). Vx and MLogP were
serious confounding factor facing the development of tau generated based on the chemical structure of the radiotracer
radiotracers is off-target brain binding, which might affect using commercial software, dproperties (Talete, Italy). KD
the quantitative analysis of the PET images as observed in values were extracted from the literature, measured via
[11C]PBB3, [18F]THK5351, and [18F]flortaucipir (also binding assays, using synthetic tau or human brain ho-
known as [18F]AV1451 or [18F]T807) [13–15]. [18F]MK6240 mogenates. MLogP was used to derive the free fractions of
Amyloid biomathematical screening methodology
Generation of In silico In vitro/model Fixed

physicochemical and Lipophilicity (MLogP) Plasma free fraction (fp) Dissociation constant Available binding sites
pharmacological Molecular volume (Vx) Tissue free fraction ( fND) (KD) density (Bavail)
parameters
Derivation of
Influx rate constant Efflux rate constant Binding potential
1TCM kinetic
parameters K1 = F(MLogP, Vx) k2 = G(K1, fp, fND) BPND = H(fND, Bavail, KD)
Plasma input function

IF
Population variation Biomathematical model Simulated TACs and SUVRs

40
K1
concentration (kBq/m)
K1 (10%) 1000 35
CTarget 30
k2 (20%) times
Ratioactivity
Simulations of 25
Bavail_HC (10%) K2/(1 + BPND) 20
population IF
Bavail_MCI (35%) 15
TACs and SUVRs K1 10
Bavail_AD (35%)
CReference 5
Monte carlo noise simulation 0
Kt 0 10 20 30 40 50 60 70 80 90
1000 SUVRs for HC, MCI and AD conditions
AUROC (Az) Effect size (Es) SUVR ratio (Sr)

1
0.8 SUVRMCI
Sensitivity
Tracer evaluation 0.6 HC-MCI CUI

SUVRHC
using CUI 0.4 (Az × Es × Sr)
MCI-AD
0.2 SUVRAD
0
0 0.2 0.4 0.6 0.8 1 SUVRMCI
1 – specificity
Figure 1: Overview of amyloid biomathematical screening methodology.
the radiotracer in tissues (fND, unitless) and in plasma (fP, 2

−2.544log(V1/3
x )−2.525
P � 10−0.121(MLogP −2.298) . (3)
unitless) from the following relationships [4]:
fND � 7.717e−1.634·MLogP , The efflux rate constant (k2, min−1) can be derived using
(1) K1, fP, and fND at equilibrium:
fP � 0.936 · f0.600
ND .
fND
k2 � · K1 . (4)
The list of 22 tau radiotracers and their respective in- fP
puts are shown in Table 1. The KD values that were utilized
for simulations are given in bold for human brain ho- The in vivo nondisplaceable binding potential (BPND,
mogenates, and italicized for synthetic tau, if available for unitless) was determined using Mintun’s equation with
comparison. Bavail, fND, and KD:
Bavail
BPND � fND · . (5)
KD
2.3. Derivation of 1TCM Kinetic Parameters. The influx rate
constant (K1, mL/cm3/min) was derived using the modified The available tau-binding sites (Bavail, nM) were mea-
Renkin and Crone equation, using compound-specific sured using enzyme-linked immunosorbent assay (ELISA).
permeability (P, cm/min), with fixed values of capillary The total amount of tau fibrils (Bavail, nM) in the frontal
surface area (S � 150 cm2/cm3 of brain) and perfusion (f � lobes, parietal lobes, and hippocampus in HC and AD were
0.6 mL/cm3/min) as follows [4, 6]: 1.5 and 16.0 nM, respectively [29], assuming a tau molecular
K1 � f􏼐1 − e−PS/f 􏼑. (2) weight of 78,928 Da (https://www.phosphosite.org).
The compound-specific permeability was derived from

the simplified Lanevskij’s permeability model, with MLogP 2.4. Simulations of Population Time Activity Curves (TACs)
and Vx as inputs [4, 6]: and SUVRs. The predicted K1, k2, and BPND were used to
Table 1: In silico MLogP and Vx and in vitro KD of 22 tau-related PET radiotracers. KD values employed for simulations are given in bold
(measured using brain homogenates) and italicized (measured using synthetic tau).
Radiotracers MLogP Vx KD References for KD
1.67 α [17]
[18F]THK523 3.19 2.11 1.99α [18]
86.5 [19]
1.45α
[18F]THK5105 3.08 2.59 [19]
2.63
[18F]THK5116 2.62 2.31 106& [12]
[18F]THK5117 2.85 2.45 2.65$ [20]
[18F]THK5125 3.08 2.59 10.2 [12]
[18F]THK5129 2.48 2.55 3.14 [12]
[18F]THK5151 2.25 2.41 7.07 [12]
[18F]THK5287 1.94 2.55 2.60 [12]
[18F]THK5307 1.71 2.41 5.60 [12]
[18F]THK5317 2.85 2.45 9.40& [21]
[18F]THK5351 2.25 2.41 2.90 [15]
[18F]THK5451 2.25 2.41 28.0 [12]
[18F]flortaucipir 1.95 1.86 14.6# [22]
[18F]T808 3.64 2.23 22.0# [11]
2.50α [10]
[11C]PBB3 2.34 2.31
6.30 [23]
18
[ F]FDDNP 2.89 2.31 36.7 α [18]
[18F]FPPDB 2.87 3.15 44.8 [24]
[11C]NML 1.98 2.51 0.700α [25]
3.30α [25]
[18F]Lansoprazol 1.75 2.37
>3998δ [11]
13.4
[26]
[11C]Astemizole 4.63 3.56 1.86α
>3998δ [11]
[18F]MK6240 2.49 1.96 0.260β [27]
[18F]JNJ64349311 ([18F]JNJ311) 2.07 1.83 7.90δ [28]
Units: MLogP (unitless), Vx (cm3/mol/100), KD (nM). $Averaged KD values (2.2, 3.1) for tau in AD brain homogenates of temporal and hippocampus.
β
Averaged KD values 0.14, 0.30, 0.25, 0.24, and 0.38 for tau in AD brain homogenates of frontal and entorhinal cortex of 5 AD. αKD values are measured using
synthetic tau (K18Δ280K) &Ki values measured using AD brain homogenates with THK5105 as competitor δKi values measured using AD brain homogenates
with T808 as competitor. #KD values measured using AD brain via autoradiography.
simulate the TACs in the target regions of HC, MCI, and AD to the mean value [29]. The amount of soluble tau in HC, MCI,
and a reference region, with a fixed arterial input function and AD was reported, but since they did not correlate well with
(IF): the amount of phosphorylated tau, these values could not be
used [31]. In our simulations, the total amount of tau fibrils in
CTarget (t) � K1 · e−k2 /(1+BPND )·t ⊗ IF(t),
(6) MCI was assumed to be the mean of that in HC and AD, with
CReference (t) � K1 · e−k2 ·t ⊗ IF(t). the same amount of variation of 35%, as used for the amyloid
simulations [5].
An input function with similar kinetics to that observed 1000 noisy TACs in both target and reference regions
in tau imaging with a fast uptake and washout is required to were generated by computer simulations with noise. In our
reflect tau kinetics. For our simulations, a fixed arterial input simulation, the target region refers to a brain region with
function was applied with fast kinetics that was derived by varying concentrations of phosphorylated tau depending on
averaging the metabolite-corrected arterial plasma input subject conditions (e.g., temporal lobe) and a reference is a
functions of 6 HC subjects injected with [11C]BF227 [30]. brain region devoid of phosphorylated tau (e.g., cerebellum).
The same K1 and k2 scaling factors of 1.23 and 1.15, re- 1000 SUVRs of each subject condition of HC, MCI, and AD
spectively, were introduced to account for the differences were determined from the ratio of the areas under the TACs
between the predicted and in vivo values [5]. The scaling factor of the target regions in HC, MCI, and AD and that of the
of BPND was modified from 0.39 to 1.0 because there were few reference region within a chosen time window. For our
reported values to determine the appropriate scaling factor. simulations, a default time window of 90–110 min was se-
Monte Carlo simulations were applied to generate 1000 TACs lected as the predicted TACs of HC, MCI, and AD appeared
in both target and reference regions with 3% noise, to reflect to reach a quasi-steady-state in this time window for almost
the noise in PETdata, and the population variation, by varying all 9 clinically applied tau radiotracers (Supplementary 2). To
K1 and k2 by 10% and 20%, respectively [5, 6]. The variations in evaluate the efficacy of fixed time windows, SUVRs were also
the tau fibrils in HC and AD were determined as 10% and 35%, determined using the literature-reported time windows for
respectively, using the ratio of the summed standard deviation the 9 clinically applied radiotracers.
2.5. Tracer Evaluation Using CUI. Az, Es, and Sr are the area 0.115, with a difference of 21.7% [38]. However, unlike [18F]
under the receiver operating characteristics curve, effect size, flortaucipir where both K1 and k2 values were determined
and SUVR ratios, respectively. The 1000 SUVR simulated using the two-tissue-compartment model with a variable
under the subject conditions of HC, MCI, and AD were used fraction [36], the reported k2 value of [18F]THK5351 was an
to determine Az, Es, and Sr for conditions-pairs of HC-MCI apparent rate constant from reference region to plasma,
and MCI-AD. CUI was then derived from the product of the which was determined using the simplified reference tissue
averaged Az (Az), Es (Es), and Sr (Sr) of conditions-pairs of model (SRTM) [38].
HC-MCI and MCI-AD with equal weightage applied: The predicted k2 value of [18F]THK5317 of 0.087 was
CUI � Az × Es × Sr. (7) close to the literature-reported value of 0.09, even though
K1 of [18F]THK5317 value differed of [18F]THK5317 from
The simulated TACs and the predicted SUVR were the clinically observed value with a difference of −39% [39].
compared to the clinical data of 9 clinically applied tau The predicted BPND values of 0.125 and 8.13 were very
radiotracers. The predicted K1, k2 and BPND values were different from the clinically observed values of 0.60 and
compared with the clinically observed values where appli- 5.11 in AD for [18F]THK5317 [39] and [18F]MK6240 [16].
cable. Finally, the list of 22 tau radiotracers (Table 1) was The predicted BPND value was fairly close to that of [11C]
evaluated using CUI. We previously developed a MATLAB- PBB3 [10]. The predicted K1 of [18F]MK6240 was close to
based program, RSwCUI, (Ver. 2014b, The MathWorks, US) the clinically observed K1 value with 2.50% difference but
[5], to support the screening of amyloid radiotracers based the predicted k2 value yielded greater difference of about
on the proposed amyloid biomathematical screening 40% [16].
methodology. The program was used for the evaluation of Table 3 shows the predicted SUVR values obtained using
tau radiotracers in this study. the default time window and literature-reported time
window of 90–110 min, and the clinically observed SUVR for
3. Results 10, 10, and 9 clinically applied tau radiotracers. The dif-
ferences in the SUVRs predicted using both time windows
Figure 2 shows the simulated TACs for the target regions of were very small for both HC and AD. The predicted SUVR
HC, MCI, and AD and reference regions of 9 clinically for HC was always greater than 1.0, but the clinically ob-
applied tau radiotracers. In general, the clinically observed served SUVR values were less than 1.0 for some radiotracers.
TACs of THK compounds of the reference region had higher In general, the clinically observed SUVR for HC and AD
peaks and faster washout in the cerebellum than the target were greater than the predicted SUVR determined using the
regions [15,32–35], while the peaks of the simulated TACs of literature-reported time window, except for [11C]PBB3 and
the reference region were always lower than that of the target [18F]MK6240, where the predicted SUVR for HC and AD
regions (Figures 2(a)–2(e)). The simulated TACs of [11C] were greater.
PBB3 (Figure 2(f )) were close to that observed clinically in The correlations between the predicted and highest
AD in the nonbinding and low-, middle-, and high-binding clinically observed SUVR for AD were similar with co-
regions [10]. The simulated TACs of [18F]flortaucipir efficients of determination, R2 of 0.90 and 0.89, re-
(Figure 2(g)) had slightly sharper peaks and faster washout spectively, using the literature-reported time window
compared to the clinically observed TACs for both HC and and the default time window (Figure 3). However, the
AD [36]. Unlike the THK compounds, the peaks of the good correlation was driven by [18F]MK6240, which had
clinically observed TACs of the target regions of [18F]flor- the highest predicted and clinically observed SUVR. Poor
taucipir were higher than that of the reference region, which correlation was observed after removing [18F]THK5351
was also observed in the simulated TACs [36]. The predicted and [18 F]MK6240. The small difference between the
TACs of [18F]T808 for both the reference and the target predicted SUVR using the default and clinical-reported
regions of HC, MCI, and AD conditions completely over- time window, and the value of R2, showed that the default
lapped with each other (Figure 2(h)). The clinically observed time window of 90–110 min was suitable for predicting
TACs of [18F]T808 appeared close to that of [18F]flortaucipir, the SUVR of the tau radiotracers (Figure 3).
but with smaller differences between the subject conditions. The simulated SUVR distribution of [18F]THK523
However, the simulated TACs showed complete overlapped across HC, MCI, and AD conditions substantially
between the HC and AD conditions with a slower uptake overlapped each other (Figure 4(a)). However, the
and washout [37]. The predicted TACs of both target and clinically observed SUVR distribution of [18F]THK523
reference regions of [18F]MK6240 showed similar fast uptake differed across different regions of interest, with HC−
but slower washout than clinically observed TACs [16]. (PIB-negative) having the smallest spread and smallest
Table 2 compares the predicted and clinically-reported values, HC+ (PIB-positive) having a relatively large
values of K1, k2, and BPND of five clinically applied tau spread and values ranging between that of HC− and AD,
radiotracers with reported kinetic parameters. For [18F] and AD subjects having the largest values and a nearly
flortaucipir, the predicted K1 and k2 values of 0.256 and similar spread as HC+ [30]. For [11C]PBB3, [18 F]
0.199, respectively, were relatively close to the reported THK5117, and [18 F]flortaucipir, the clinically observed
averaged cerebellar K1 and k2 values of 0.26 and 0.17, re- SUVR distributions were generally larger for AD than
spectively [36]. The predicted k2 value of [18F]THK5351 was HC for all regions of interest analyzed, in terms of the
0.140, which was higher than the clinically observed value of spread and absolute values [14, 34, 40]. The trend of the
25 25 25
Radioactivity concentration
20 20 20
(kBq/mL)
(kBq/mL)
(kBq/mL)
15 15 15
10 10 10
5 5 5
0 0 0
0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120
Time (min) Time (min) Time (min)
Reference Target (MCI) Reference Target (MCI) Reference Target (MCI)

Target (HC) Target (AD) Target (HC) Target (AD) Target (HC) Target (AD)
(a) (b) (c)

25 25 25
20 20 20
(kBq/mL)
(kBq/mL)
(kBq/mL)
15 15 15
10 10 10
5 5 5
0 0 0
0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120

(d) (e) (f )
25 25 60
20 20 50
40
(kBq/mL)
(kBq/mL)
(kBq/mL)
15 15
30
10 10
20
5 5 10
0 0 0
0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120

(g) (h) (i)
Figure 2: Simulated TACs of target regions of HC, MCI, and AD and reference regions for (a) [18F]THK523, (b) [18F]THK5105, (c) [18F]
THK5117, (d) [18F]THK5317, (e) [18F]THK5351, (f ) [11C]PBB3, (g) [18F]flortaucipir, (h) [18F]T808, and (i) [18F]MK6240 from 0–120 min.
simulated SUVR population distribution was close to KD values measured using synthetic tau were smaller (Ta-
that observed clinically for HC and AD conditions ble 1). The ranking of the CUI values generated using the KD
(Figures 4(b)–4(d)). This supported the use of 35%, 35% values measured with synthetic tau and brain homogenates
and 10% variations in Bavail for population simulations. differed for [18F]THK523, [18F]THK5105, and [11C]PBB3. 10
Figure 5 shows the CUI distribution of 22 tau-related out of 16 tau radiotracers had CUI values higher than the
radiotracers. Among the clinically applied tau radiotracers, recommend CUI value of 3.0, where the results were simu-
[18F]MK6240 was ranked first, followed by [18F]THK5351, lated using KD values measured with human brain homog-
[18F]THK5117, [11C]PBB3, [18F]flortaucipir, [18F]THK5317, enates. Apart from [18F]THK523, [18F]THK5317, [18F]T808,
[18F]FDDNP, [18F]T808, and [18F]THK523, based on the KD and [18F]FDDNP, the other 6 clinically applied tau radio-
values measured using AD brain homogenates. For candi- tracers yielded high CUI values. The CUI values ranged from
date radiotracers, [18F]THK5287 was ranked first based on about 0.0 to 16.2, which ranged wider than that for amyloid.
the KD values measured using AD brain homogenates, while
[11C]NML was ranked first based on the KD values measured 4. Discussion
using heparin-induced tau polymer (HITP) (Table 1). The
CUI values generated using the KD values for the synthetic In this paper, we evaluated the feasibility of extending a
tau were higher than those of the brain homogenates as the previously developed amyloid biomathematical screening
Table 2: Comparison of predicted and clinically observed K1, k2, and BPND values of four clinically applied tau radiotracers.
Literature
Predicted %
Radiotracers Clinically observed
Parameters Region References values diff
values
Cerebellum excluding
K1 0.26 0.256 −1.54
[18F]flortaucipir vermis [36]
k2 0.17 0.199 17.1
[18F]THK5351 (S-enantiomer of [18F]
k2′‡ Target β
0.115 [38] 0.140 21.7
THK5151)
δ
K1 Target 0.33 0.202 −38.8
[35]
[18F]THK5317 (S-enantiomer of [18F] k2 0.09 0.087 −3.33
THK5117) BPND
Putamen 0.60 [39] 0.125 −79.2
(AD)∗
BPND High-binding cortical
[11C]PBB3 0.37 [10] 0.427 15.4
(AD)¶ regions
K1 0.246 0.252 2.50
Posterior cingulate
[18F]MK6240 k2 0.099 [16] 0.138 39.2
cortex
BPND§ 5.11 8.13 59.2
β
Target ROIs: anterior cingulate, brainstem, caudate nucleus, eroded white matter, entorhinal cortex, frontal cortex, fusiform gyrus, hippocampus, inferior
temporal cortex, lingual gyrus, middle temporal gyrus, occipital cortex, pallidum, parahippocampal gyrus, parietal cortex, posterior cingulate, precuneus,
putamen, thalamus. δTarget ROIs: thalamus, putamen, hippocampus, amygdala, parietal cortex, frontal cortex, sensory motor cortex, occipital cortex,
midbrain, entorhinal cortex, and temporal cortex. ∗ BPND � DVR-1, where DVR was determined using reference Logan, averaged from 4 prodromal AD.
¶
BPND determined using MRTM0. §BPND determined using k3/k4 using 2T4CM in 7 symptomatic individuals classified as MCI and AD. ‡k2′ optimized from
fitting all target ROIs using SRTM with cerebellum as the reference region.
Table 3: Comparison of predicted (literature-reported and default time window of 90–110 min) and clinically observed SUVR (highest
SUVR in AD) of HC and AD conditions.
Predicted SUVR Clinically observed SUVR
Highest in
Clinically applied radiotracers Default Literature
AD Regions Time window (min) References
HC AD HC AD HC AD
[18F]THK523 1.00 1.01 1.00 1.01 0.96 1.81 ITL 60–90 [32]
[18F]THK5105 1.03 1.34 1.03 1.35 1.41 1.52 PU 90–100 [33]
[18F]THK5117 1.04 1.47 1.05 1.56 1.57 1.77 PU 50–60 [34]
[18F]THK5317 1.01 1.13 1.01 1.14 — — — — —
[18F]THK5351 1.10 2.11 1.11 2.38 2.14 2.98 HIP 50–60 [15]
[18F]flortaucipir 1.03 1.35 1.03 1.35 1.17 2.19 ITL 80–100 [40]
[18F]T808 1.00 1.02 1.00 1.02 0.94 1.52 LTL 80–100 [35]
[11C]PBB3 1.04 1.43 1.05 1.55 0.85 1.42 Global# 30–50 [10]
[18F]FDDNP 1.00 1.03 1.00 1.04 1.24 1.37 ACG 45–55 [41]
[18F]MK6240 1.78 9.94 1.78 9.93 — ∼5∗ PRE 90–110 [16]
ITL � inferior temporal lobe, LTL � lateral temporal lobe, PU � putamen, PAR � parietal lobe, HIP � hippocampus, ACG � anterior posterior cingulate, PRE
� precuneus. #Global � cerebral cortex for HC and high binding ROI for AD. ∗ SUVR is approximated from the plot, taking the highest SUVR in AD.
methodology to support the screening of candidate tau [18F]flortaucipir, and [11C]PBB3 (Figure 4). Both the pre-
radiotracers during compound development. 22 clinically dicted and clinically observed SUVR values were less than
applied and candidate tau-related radiotracers were thus 1.0 in HC for some radioligands, especially those with a
used to investigate the CUI ranking of clinically applied and lower selectivity for tau (e.g., [18F]THK523). The clinically
candidate tau radiotracers. observed SUVR of AD is much higher than that of HC.
However, there is little difference in the predicted SUVR for
[18F]THK523. This shows that the predictions were less
4.1. Comparison of Simulated TACs and SUVR Distribution. accurate for tau compounds with a lower selectivity for the
The simulated TACs were very different from the clinically target. Poor predictions might be due to binding to other
observed TACs of [18F]THK523 and [18F]T808, but were β-sheet structured proteins or off-target sites shown in the
only slightly different for that of [18F]THK5117, [18F] clinical data, whereas the predicted values showed the
THK5351, [18F]flortaucipir, [11C]PBB3, and [18F]MK6240 binding of the radiotracers to only the target site. Non-
(Figure 2). The simulated SUVR distributions were different specific binding in white matter may also lead to spill-over
for [18F]THK523 but were similar to the clinically observed into the surrounding cortical regions, leading to higher
results under HC and AD conditions for [18F]THK5117, clinically observed SUVRs. The issue of non-specific binding
11
10 Highest SUVR (default, x--): [18F]MK6240
9 y = 2.31x – 2.72, R2 = 0.88
Predicted SUVR (AD)

8 Highest SUVR (literature, -):
7 y = 2.31x – 2.67, R2 = 0.89
6
5
4
3
2 [18F]THK5351
1
0
0 1 2 3 4 5 6
Clinically observed SUVR (AD)
Figure 3: Correlations of clinically observed and predicted SUVR values using literature-stated time window (o-) and default time window
(x--).
1.08 1.8
1.06 1.7
1.04 1.6
Simulated SUVR
Simulated SUVR
1.5
1.02
1.4
1.00
1.3
0.98 1.2
0.96 1.1
0.94 1.0
HC MCI AD HC MCI AD
Representative subject conditions Representative subject conditions
(a) (b)
1.6 1.7
1.5 1.6
Simulated SUVR
Simulated SUVR
1.4 1.5
1.3 1.4
1.3
1.2
1.2
1.1
1.1
1.0
1.0
HC MCI AD HC MCI AD
Representative subject conditions Representative subject conditions
(c) (d)
Figure 4: Simulated SUVR distributions of (a) [ F]THK523, (b) [ F]THK5117, (c) [18F]flortaucipir, (d) [11C]PBB3.
18 18
is more apparent for tau radiotracers with lower tau-binding S-enantiomers of [18F]THK5151 and [18F]THK5117 re-
selectivity, such as [18F]THK523 and [18F]THK5117 spectively (Table 2). The predictions for BPND were generally
(Table 2). poor for the three clinically-reported tau radiotracers (Ta-
ble 2). This may be due to the use of a simplified 1TCM for
prediction, even though 2TCM was reported to be more
4.2. Comparison of Predicted 1TCM and SUVR. The pre- suitable for many clinically applied tau radiotracers. The
diction for the K1 and k2 values of the tau radiotracers simplified 1TCM was selected even though 2TCM is more
appeared to work well in racemic compounds (e.g., [18F] accurate for modeling tau kinetics as the prediction of a
flortaucipir), but not as well for enantiomeric compounds larger number of microparameters may be difficult to es-
like [18F]THK5351 and [18F]THK5317, which are timate reliably. Moreover, the 1TCM worked reasonably
18
16
14
12
CUI
10
0
[11C]Astemizole
[18F]THK523
[11C]Astemizole∗
[18F]T808
[18F]THK5116
[18F]FPPDB
[18F]FPDNP
[18F]THK5125
[18F]THK5451
[18F]THK5317
[18F]THK5105
[18F]flortaucipir
[11C]PBB3
[18F]THK523∗
[18F]THK5117
[18F]THK5151
[18F]JNJ64349311
[18F]THK5105∗
[18F]THK5129
[11C]PBB3∗
[18F]THK5351
[18F]THK5307
[18F]THK5287
[18F]Lansoprazol∗
[11C]NML∗
[18F]MK6240
KD measured using synthetic tau
KD measured using AD brain homogenates
Ki measured with competitors
Figure 5: CUI distributions of 22 tau-related PET radiotracers.
well in predicting the kinetics of the amyloid radiotracers, 4.3. Comparison of Tau Radiotracers with CUI. The CUI
even though 2TCM was reported to be more suitable [5]. value of [18F]flortaucipir was large while the CUI value of
Other possible reasons for the poorer BPND predictions in- [18F]T808 was very small and does not appear to be a
cluded differences in binding to the plasma proteins due to promising clinical tau radiotracer. Similarly, [18F]THK523
the enantiomeric properties of the radiotracers [42], me- also yielded a small CUI value, even though studies showed
tabolites crossing the BBB for [11C]PBB3 [10], binding of tau that it could be applied clinically. [18F]THK523, [18F]Lan-
radiotracers to other similar β-sheet structures (Aβ and α-- soprazole, and [11C]Astemizole yielded small CUI values
synuclein), or off-target binding in target regions of interest using the KD values measured using human brain homog-
[13–15]. The predicted 1TCM parameters and SUVR, as well enates, which differed greatly from that measured using
as the simulated TACs and SUVR distribution, were com- synthetic tau. KD or Ki values measured using AD brain
pared to clinically observed data where applicable. However, homogenates were very different from those measured using
we were limited by the small number of reported kinetic heparin-induced tau polymer (HITP) (Table 1). This is
parameters and SUVR values to fully assess the amyloid because HITP is composed of only 3R and/or 4R, and hence
biomathematical model for screening tau radiotracers. may not undergo the same phosphorylation process as
The predicted and highest clinicallyobserved SUVR data human tau [19, 43]. On the other hand, the KD or Ki values of
for AD correlated well using fixed time window of 90– amyloid radiotracers measured using synthetic tau and AD
110 min and the literature-reported time window with R2 brain homogenates did not differ greatly [5]. The huge
values of 0.88 and 0.89 respectively, for 9 clinically applied difference in the KD values measured using human brain
tau radiotracers (Figure 3). However, the results were driven homogenates and synthetic tau were much greater for [18F]
mostly by [18F]MK6240. Some of the clinically applied tau THK523 than for [18F]THK5105 (Table 1). This might also
radiotracers ([18F]THK523, [18F]THK5351 and [18F]flor- indicate the binding preferences of [18F]THK523 to certain
taucipir) did not have high selectivity for tau, which may tau-binding sites available on synthetic tau, that were fewer
have contributed to smaller predicted values as the predicted in numbers in human brain homogenates. Therefore, it is
values were based on binding to a single target site but the important to determine the binding affinity of tau radio-
off-target binding or specific binding to other β-sheet tracers to different subtypes of tau protein and other β-sheet
structures (e.g., amyloid) may yield higher clinical SUVR structures such as Aβ and α-synuclein.
values. The predicted TACs of [18F]T808 exhibited a much [18F]THK5351 yielded higher clinically observed SUVR
slower clearance compared to the clinically observed ki- than [18F]THK5117 in the same AD patients, with lower
netics, which resulted in a large difference between the white matter binding [15]. [18F]THK5351 was also reported
predicted and clinically observed SUVR. This may be due to to have a higher signal-to-noise ratio (SNR), and a lower
the poor predictive ability of in silico parameters for [18F] non-specific binding in white matter than [18F]THK5105
T808, which has a unique chemical structure. and [18F]THK5117 [8]. Similarly, the CUI value of [18F]
THK5351 was higher than [18F]THK5105 and [18F] methodology could not be used to predict the possibility of
THK5117. [11C]PBB3, [18F]flortaucipir, and [18F]THK5105 metabolites crossing this barrier.
have nearly similar CUI values (Figure 5), but the difference Off-target binding was observed in some clinically ap-
in the clinically observed SUVR values between HC and AD plied tau radiotracers. [18F]flortaucipir was reported to show
were greatest in [18F]flortaucipir, followed by [11C]PBB3 specific binding in the midbrain, vessels, iron-associated
then [18F]THK5105 (Table 3). This difference may be at- regions (e.g., basal ganglia), substantia nigra, calcifications
tributed to the tau subtypes that [11C]PBB3 is binding. [18F] in the choroid plexus, and leptomeningeal melanin [13].
THK5351 and [18F]flortaucipir was reported to bind to the [11C]PBB3 was reported to accumulate in the venous sinuses,
same targets but with different affinities, while [11C]PBB3 basal ganglia, and thalamus, while its fluorinated com-
seems to bind to a different tau subtype [44]. If the tau pounds showed off-target binding in the choroid plexus
subtype that [11C]PBB3 binds to is of a lower concentration [14, 44]. [18F]THK5351 was reported to bind to monoamine
in subject, the clinical SUVR will become smaller. The oxidase B (MAO-B), which is highly expressed throughout
difference between the clinically observed results and CUI the brain, and thus, its tau binding data needs to be corrected
ranking showed that the evaluation of the clinical usefulness for MAO-B binding [47]. [18F]MK6240 was reported to have
of tau radiotracer based on binding to a single target could reduced off-target binding on the whole but showed off-
not reflect the actual in vivo binding in subjects. High tau target binging in regions such as the retina, substantia nigra,
selectivity and off-target binding affect the comparison of the ethmoid sinus, and dura matter [16]. Depending on the
in vivo binding of tau radiotracers, which are less prominent region of off-target binding, the effects may not limit PET
in amyloid radiotracers. Despite the differences in CUI quantification due to little or no anatomical overlap of the
rankings, the clinically applied tau radiotracers had CUI target regions of interest (ROIs) with off-target regions.
values above the recommended value especially for those Accurate PET quantification is also less affected if the ra-
with high SNR. Thus, the screening methodology can still diotracer has high target selectivity or if the concentrations
provide confidence in the decision-making of moving of the off-target binding sites are much lower compared to
candidate radiotracers for clinical studies. that of the target [48]. Off-target binding may be one of the
contributing factors that led to the observed differences
between simulation and the clinical data of tau PET ra-
4.4. Limitations of Screening Methodology. Few measure- diotracers. The possibility of binding to off-targets is difficult
ments of tau concentration in postmortem human brains to predict, and systematic screening is required to determine
using ELISA have been reported, and these values are very the binding of the candidate compound to a wide range of
different [17, 29, 45, 46]. In addition, these reported tau proteins. This will increase the time and cost of compound
concentrations were mostly measured using normal-aged screening. The amyloid biomathematical screening meth-
control and AD brains, with very little data on the tau odology could not predict off-target binding, and the in-
concentration in MCI. As such, the simulated SUVR dis- clusion of multiple binding sites appeared to be required for
tribution might not reflect the clinically observed MCI re- tau radiotracers to correct for this issue.
sult. Moreover, the input function of the amyloid radiotracer
[11C]BF227 was used for simulations. Thus far, the input
functions of only three clinically applied tau radiotracers of 4.5. Feasibility of Extending to the Screening of Tau
[11C]PBB3 [10], [18F]flortaucipir [36], and [18F]MK6240 [16] Radiotracers. To date, the comparison of multiple tau ra-
have been reported. The arterial input functions of these diotracers has been performed via in vitro competition
radiotracers were similar in HC and AD, with a fast uptake binding assays in human brain sections, using human AD
and a fast washout, and the shape of the curves was similar to brain homogenates [11, 12] or by means of preclinical
that of [11C]BF227 as used in the simulation. Although the imaging [38]. There is a lack of consideration of the possible
shape of the input function of these two radiotracers was in vivo kinetics of the radiotracers during development,
similar to that of [11C]BF227, the shape of the arterial input which may lead to poor clinical performance [4–6]. The use
function might be different for other tau radiotracers. Thus, of in silico data can support predictions of tracer kinetics and
we evaluated the effect of the input function on the outcome increases confidence in clinical translation, in addition to
using four different input functions with fast kinetics for HC facilitating radiotracer comparisons. The weak SUVR cor-
and AD subjects injected with [11C]BF227 or [18F]FACT, relation was obtained between the predicted and clinically
with areas under the input function curves from 0 to 120 min observed SUVR results, mostly due to the small SUVR values
of 536 (default), 649, 434, and 306 (kBq/mL) min. The % for tau radiotracers with poorer tau selectivity. However,
COV of the predicted SUVR was less than 7.0 for all con- there are very few reported kinetic parameters to assess the
ditions and radiotracers, while %COV of the CUI was less limitations of the screening methodology. The TACs, SUVR
than 7.0 for all except the poor radiotracers, namely, [18F] distribution, and CUI rankings differed primarily for tau
FDDNP, [18F]FPPDB, and [11C]Astemizole. This showed radiotracers with low selectivity to tau. This showed that the
that the results would not be changed significantly using evaluation of the clinical usefulness of tau radiotracer based
input functions with similar kinetics. However, there were on binding to a single target could not fully reflect the actual
also issues with metabolites crossing the BBB (e.g., [11C] in vivo binding in subjects since they also exhibited binding
PBB3), but the amyloid biomathematical screening preferences to nontarget sites. Thus, it is not feasible to
directly apply the amyloid biomathematical screening Supplementary Materials

methodology to tau radiotracers due to the increased
complexity of evaluating the binding of tau radiotracers, Supplementary 1. Table 1: Predicted K1, k2, and BPND values
namely, target-binding, off-target binding, and non-specific in HC and AD of 9 clinically applied tau-related radiotracers.
binding. More work is required to improve the accuracy of Supplementary 2. Figure 1: Quasi-steady-state of TACs of
predicting the clinical usefulness of tau radiotracers by in- HC (gray), MCI (blue), and AD (orange) conditions of 9
cluding possible binding to other β-sheet structures or off- clinically applied tau radiotracers from 0 to 120 min. The
target sites. However, the high CUI values generated for box highlighted the time window of 90 to 110 min, which
clinically applied tau radiotracers with high SNR showed was applied as the default time window for simulating
that the screening methodology could be used to increase SUVRs.
confidence in decision-making when choosing candidate
radiotracers for further evaluation.
References
5. Conclusions [1] R. J. Perrin, A. M. Fagan, and D. M. Holtzman, “Multimodal
techniques for diagnosis and prognosis of Alzheimer’s dis-
The predicted TACs, SUVR, and CUI ranking differed for ease,” Nature, vol. 461, no. 7266, pp. 916–922, 2009.
some clinically applied tau radiotracers, especially those [2] C. R. Jack, D. A. Bennett, K. Blennow et al., “NIA-AA Re-
with lower selectivity for tau. This showed that the eval- search Framework: toward a biological definition of Alz-
uation of the clinical usefulness of tau radiotracer based on heimer’s disease,” Alzheimer’s and Dementia, vol. 14, no. 4,
binding to a single target could not reflect the actual in vivo pp. 535–562, 2018.
[3] R. C. Garner and G. Lappin, “Commentary,” British Journal of
tau binding in subjects due to more challenges in evaluating
Clinical Pharmacology, vol. 61, no. 4, pp. 367–370, 2006.
the in vivo binding of tau radiotracers, such as off-target
[4] Y. Arakawa, Y. Nai, M. Shidahara et al., “Prediction of the
binding and high tau selectivity, compared to amyloid clinical SUV ratio in amyloid PET imaging using a bio-
radiotracers. The inclusion of possible binding to other mathematic modeling approach toward the efficient devel-
β-sheet structures or off-target sites and the binding af- opment of a radioligand,” Journal of Nuclear Medicine, vol. 58,
finities to different target sites would improve the accuracy no. 8, pp. 1285–1292, 2017.
of the prediction. From our results, clinically applied tau [5] Y. H. Nai, M. Shidahara, C. Seki, and H. Watabe, “Bio-
radiotracers with higher SNR, such as [18F]MK6240 and mathematical screening of amyloid radiotracers with clinical
[18F]THK5351, had higher CUI rankings. This supported usefulness index,” Alzheimer’s and Dementia: Translational
the use of the screening methodology in radiotracer de- Research & Clinical Interventions, vol. 3, no. 4, pp. 542–552,
velopment by allowing comparison of candidate radio- 2017.
tracers with clinically-applied radiotracers based on SUVR, [6] Q. Guo, M. Brady, and R. N. Gunn, “A biomathematical
with respect to binding to a single target. Our results will modeling approach to central nervous system radioligand
discovery and development,” Journal of Nuclear Medicine,
hopefully provide some insights to guide the development
vol. 50, no. 10, pp. 1715–1723, 2009.
of in silico models in supporting the development of tau
[7] L. Buée, T. Bussière, V. Buée-Scherrer, A. Delacourte, and
radiotracers. P. R. Hof, “Tau protein isoforms, phosphorylation and role in
neurodegenerative disorders11These authors contributed
equally to this work,” Brain Research Reviews, vol. 33, no. 1,
Data Availability
pp. 95–130, 2000.
The program (RSwCUI) used for TACs simulation and CUI [8] V. L. Villemagne, M. T. Fodero-Tavoletti, C. L. Masters, and
evaluation can be download from http://www.rim.cyric. C. C. Rowe, “Tau imaging: early progress and future di-
rections,” Lancet Neurology, vol. 14, no. 1, pp. 114–124, 2015.
tohoku.ac.jp/software/CUI-Software. The predicted K1, k2,
[9] H. Hashimoto, K. Kawamura, N. Igarashi et al., “Radiosyn-
and BPND values in HC and AD of 9 clinically applied tau- thesis, photoisomerization, biodistribution, and metabolite
related radiotracers are provided in the interest of readers analysis of 11C-PBB3 as a clinically useful PET probe for
and are included within the supplementary information imaging of tau pathology,” Journal of Nuclear Medicine,
file. vol. 55, no. 9, pp. 1532–1538, 2014.
[10] Y. Kimura, M. Ichise, H. Ito et al., “PET quantification of tau
Conflicts of Interest pathology in human brain with 11C-PBB3,” Journal of Nu-
clear Medicine, vol. 56, no. 9, pp. 1359–1365, 2015.
The authors declare that there are no conflicts of interest [11] L. Declercq, S. Celen, J. Lecina et al., “Comparison of new tau
regarding the publication of this paper. PET-tracer candidates with [18F]T808 and [18F]T807,” Mo-
lecular Imaging, vol. 15, article 153601211562492, 2016.
[12] T. Tago, S. Furumoto, N. Okamura et al., “Preclinical eval-
Acknowledgments uation of [18F]THK-5105 enantiomers: effects of chirality on
its effectiveness as a tau imaging radiotracer,” Molecular
This study was supported by Grants-in-Aid for Scientific Imaging and Biology, vol. 18, no. 2, pp. 258–266, 2015.
Research (B) (no. 17H04118) from the Ministry of Educa- [13] V. J. Lowe, G. Curran, P. Fang et al., “An autoradiographic
tion, Culture, Sports, Science and Technology (MEXT), evaluation of AV-1451 Tau PET in dementia,” Acta Neuro-
Japanese Government. pathologica Communications, vol. 4, no. 1, 2016.
[14] M. Maruyama, H. Shimada, T. Suhara et al., “Imaging of tau disease brain,” European Neurology, vol. 42, no. 4, pp. 205–
pathology in a tauopathy mouse model and in Alzheimer 210, 1999.
patients compared to normal controls,” Neuron, vol. 79, no. 6, [30] M. Shidahara, H. Watabe, M. Tashiro et al., “Quantitative
pp. 1094–1108, 2013. kinetic analysis of PET amyloid imaging agents [11C]BF227
[15] R. Harada, N. Okamura, S. Furumoto et al., “18F-THK5351: a and [18F]FACT in human brain,” Nuclear Medicine and Bi-
novel PET radiotracer for imaging neurofibrillary pathology ology, vol. 42, no. 9, pp. 734–744, 2015.
in alzheimer disease,” Journal of Nuclear Medicine, vol. 57, [31] P. Han, G. Serrano, T. G. Beach et al., “A quantitative analysis
no. 2, pp. 208–214, 2015. of brain soluble tau and the tau secretion factor,” Journal of
[16] T. A. Pascoal, M. Shin, M. S. Kang et al., “In vivo quantifi- Neuropathology and Experimental Neurology, vol. 76, no. 1,
cation of neurofibrillary tangles with [18F]MK-6240,” Alz- pp. 44–51, 2017.
heimer’s Research and Therapy Internet, vol. 10, no. 1, 2018. [32] V. L. Villemagne, S. Furumoto, M. T. Fodero-Tavoletti et al.,
[17] M. T. Fodero-Tavoletti, N. Okamura, S. Furumoto et al., “18F- “In vivo evaluation of a novel tau imaging tracer for Alz-
THK523: a novel in vivo tau imaging ligand for Alzheimer’s heimer’s disease,” European Journal of Nuclear Medicine and
disease,” Brain, vol. 134, no. 4, pp. 1089–1100, 2011. Molecular Imaging, vol. 41, no. 5, pp. 816–826, 2014.
[18] R. Harada, N. Okamura, S. Furumoto et al., “Comparison of [33] N. Okamura, S. Furumoto, M. T. Fodero-Tavoletti et al.,
the binding characteristics of [18F]THK-523 and other am- “Non-invasive assessment of Alzheimer’s disease neurofi-
yloid imaging tracers to Alzheimer’s disease pathology,” brillary pathology using [18F]-THK5105 PET,” Brain, vol. 137,
European Journal of Nuclear Medicine and Molecular Imaging, no. 6, pp. 1762–1771, 2014.
vol. 40, no. 1, pp. 125–132, 2012. [34] R. Harada, N. Okamura, S. Furumoto et al., “[18F]THK-5117
[19] N. Okamura, S. Furumoto, R. Harada et al., “Novel 18F-labeled PET for assessing neurofibrillary pathology in Alzheimer’s
arylquinoline derivatives for noninvasive imaging of tau disease,” European Journal of Nuclear Medicine and Molecular
pathology in Alzheimer disease,” Journal of Nuclear Medicine, Imaging, vol. 42, no. 7, pp. 1052–1061, 2015.
vol. 54, no. 8, pp. 1420–1427, 2013. [35] M. Jonasson, A. Wall, K. Chiotis et al., “Tracer kinetic analysis
[20] L. Lemoine, L. Saint-Aubert, A. Marutle et al., “Visualization of (S)-18F-THK5117 as a PET tracer for assessing tau pa-
of regional tau deposits using 3H-THK5117 in Alzheimer thology,” Journal of Nuclear Medicine, vol. 57, no. 4,
brain tissue,” Acta Neuropathologica Communications, vol. 3, pp. 574–581, 2016.
no. 1, p. 40, 2015. [36] D. W. Wooten, N. J. Guehl, E. E. Verwer et al., “Pharma-
[21] T. Tago, Development of PET probes for imaging of tau pa-
cokinetic evaluation of the tau PET radiotracer 18F-T807 (18F-
thology in Alzheimer’s disease, Ph.D. Thesis, Graduate School
AV-1451) in human subjects,” Journal of Nuclear Medicine,
of Pharmaceutical Sciences, Tohoku University, Sendai, Ja-
vol. 58, no. 3, pp. 484–491, 2016.
pan, 2016.
[37] D. T. Chien, A. K. Szardenings, S. Bahri et al., “Early clinical
[22] C.-F. Xia, J. Arteaga, G. Chen et al., “[18F]T807, a novel tau
PET imaging results with the novel PHF-tau radioligand
positron emission tomography imaging agent for Alzheimer’s
[F18]-T808,” Journal of Alzheimer’s Disease, vol. 38, no. 1,
disease,” Alzheimer’s and Dementia, vol. 9, no. 6, pp. 666–676,
pp. 171–184, 2014.
2013.
[38] S. N. Lockhart, S. L. Baker, N. Okamura et al., “Dynamic PET
[23] M. Ono, N. Sahara, K. Kumata et al., “Distinct binding of PET
measures of tau accumulation in cognitively normal older
ligands PBB3 and AV-1451 to tau fibril strains in neurode-
adults and alzheimer’s disease patients measured using [18F]
generative tauopathies,” Brain, vol. 12, p. aww339, 2017.
[24] K. Matsumura, M. Ono, H. Kimura et al., “18F-Labeled THK-5351,” PLoS One, vol. 11, no. 6, Article ID e0158460,
phenyldiazenyl benzothiazole for in vivo imaging of neuro- 2016.
fibrillary tangles in alzheimer’s disease brains,” ACS Medicinal [39] K. Chiotis, L. Saint-Aubert, I. Savitcheva et al., “Imaging in-
Chemistry Letters, vol. 3, no. 1, pp. 58–62, 2011. vivo tau pathology in Alzheimer’s disease with THK5317 PET
[25] M. V. Fawaz, A. F. Brooks, M. E. Rodnick et al., “High affinity in a multimodal paradigm,” European Journal of Nuclear
radiopharmaceuticals based upon lansoprazole for PET im- Medicine and Molecular Imaging, vol. 43, no. 9, pp. 1686–
aging of aggregated tau in Alzheimer’s disease and progressive 1699, 2016.
supranuclear palsy: synthesis, preclinical evaluation, and lead [40] K. A. Johnson, A. Schultz, R. A. Betensky et al., “Tau positron
selection,” ACS Chemical Neuroscience, vol. 5, no. 8, emission tomographic imaging in aging and early Alzheimer
pp. 718–730, 2014. disease,” Annals of Neurology, vol. 79, no. 1, pp. 110–119, 2015.
[26] L. E. Rojo, J. Alzate-Morales, I. N. Saavedra, P. Davies, and [41] C. Tauber, E. Beaufils, C. Hommet et al., “Brain [18F]FDDNP
R. B. Maccioni, “Selective interaction of lansoprazole and binding and glucose metabolism in advanced elderly healthy
astemizole with tau polymers: potential new clinical use in subjects and alzheimer’s disease patients,” Journal of Alz-
diagnosis of Alzheimer’s disease,” Journal of Alzheimer’s heimer’s Disease, vol. 36, no. 2, pp. 311–320, 2013.
Disease, vol. 19, no. 2, pp. 573–589, 2010. [42] T. Itoh, Y. Saura, Y. Tsuda, and H. Yamada, “Stereoselectivity
[27] E. D. Hostetler, A. M. Walji, Z. Zeng et al., “Preclinical and enantiomer-enantiomer interactions in the binding of
characterization of 18F-MK-6240, a promising PET tracer for ibuprofen to human serum albumin,” Chirality, vol. 9, no. 7,
in vivo quantification of human neurofibrillary tangles,” pp. 643–649, 1997.
Journal of Nuclear Medicine, vol. 57, no. 10, pp. 1599–1606, [43] M. Ariza, H. C. Kolb, D. Moechars, F. Rombouts, and
2016. J. I. Andrés, “Tau positron emission tomography (PET) im-
[28] F. J. R. Rombouts, J. I. Andrés, M. Ariza et al., “Discovery of aging: past, present, and future,” Journal of Medicinal
N-(Pyridin-4-yl)-1,5-naphthyridin-2-amines as potential tau Chemistry, vol. 58, no. 11, pp. 4365–4382, 2015.
pathology PET tracers for Alzheimer’s disease,” Journal of [44] L. Lemoine, P. G. Gillberg, M. Svedberg et al., “Comparative
Medicinal Chemistry, vol. 60, no. 4, pp. 1272–1291, 2017. binding properties of the tau PET tracers THK5117,
[29] M. Herrmann, S. Golombowski, K. Kräuchi et al., “ELISA- THK5351, PBB3, and T807 in postmortem Alzheimer brains,”
quantitation of phosphorylated tau protein in the Alzheimer’s Alzheimer’s Research and Therapy, vol. 9, no. 1, p. 96, 2017.
[45] E. B. Mukaetova-Ladinska, Z. Abdel-All, E. S. Mugica et al.,

“Tau proteins in the temporal and frontal cortices in patients
with vascular dementia,” Journal of Neuropathology and
Experimental Neurology, vol. 74, no. 2, pp. 148–157, 2015.
[46] R. Hashimoto, Y. Nakamura, I. Tsujio, H. Tanimukai,
T. Kudo, and M. Takeda, “Quantitative analysis of neuro-
filament proteins in Alzheimer brain by enzyme linked im-
munosorbent assay system,” Psychiatry and Clinical
Neurosciences, vol. 53, no. 5, pp. 587–591, 1999.
[47] K. P. Ng, T. A. Pascoal, S. Mathotaarachchi et al., “Mono-
amine oxidase B inhibitor, selegiline, reduces 18F-THK5351
uptake in the human brain,” Alzheimer’s Research and
Therapy, vol. 9, no. 1, p. 25, 2017.
[48] G. C. Van de Bittner, E. L. Ricq, and J. M. Hooker, “A
philosophy for CNS radiotracer design,” Accounts of Chemical
Research, vol. 47, no. 10, pp. 3127–3134, 2014.
Hindawi
https://doi.org/10.1155/2018/1981627
Research Article
Continuous Blood Pressure Estimation Based on Two-Domain
Fusion Model
Qian Wang,1,2 Yajie Xu ,2 Guoqiang Zeng,1 and Mingshan Sun 2
1
College of Nuclear Technology and Automation Engineering, Chengdu University of Technology, Chengdu 610051,
Sichuan, China
2
Medical Imaging Department, Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences,
Suzhou 215163, Jiangsu, China
Correspondence should be addressed to Yajie Xu; xuyj@sibet.ac.cn
Received 17 September 2018; Accepted 6 November 2018; Published 16 December 2018
Copyright © 2018 Qian Wang et al. This is an open access article distributed under the Creative Commons Attribution License,
Blood pressure (BP) is one of the indispensable elements of physiological health characteristics and a significant indicator for
predicting and diagnosing hypertension and cardiovascular diseases. This paper proposes a two-domain fusion model to estimate
BP continuously from pulse wave acquired with a pressure sensor. Method. The optimal external pressure applied on the pressure
sensor is first determined in order to capture pulse wave in the radial artery. The captured pulse wave is then processed in both the
time and frequency domains via filtering and fast Fourier transform. Finally, a set of features are extracted from these two domains
and input into a neural network along with blood pressure values measured by a commercial sphygmomanometer for training.
The model is then tested on new data for accuracy evaluation. Results. The proposed two-domain fusion method achieved a high
degree of accuracy in measuring blood pressure.
1. Introduction electrocardiogram (ECG) module, which brings com-

plexity and inconvenience to portable blood pressure
According to the American Journal of Medicine, between measurement. Wu et al. only used ECG signal to estimate
1990 and 2015, the rate of blood pressure (BP) greater than BP. They took 6 features of the ECG and calculated the time
140 mmHg rose from 17307/100000 to 20526/100000 and between the relevant features. Two sets of ECG signal
the associated mortality rate from 97.9/100000 to 106.3/ containing 4 time periods were combined and input into a
100000 [1]. Admittedly, abnormal BP can cause a burden on neural network for training to get the BP. However, this
the heart, which increases the risk of cardiovascular diseases model cannot simultaneously derive the BP and is sig-
[2, 3]. The traditional way of measuring blood pressure is to nificantly different compared with the target value [6].
place a cuff on the upper arm and then detect changes in the Secondly, PTT has a significant correlation with systolic
pressure inside the cuff during inflation and deflation to blood pressure but little relationship with diastolic blood
obtain systolic blood pressure (SBP) and diastolic blood pressure [7]. Xu et al. used PTT and photoplethysmogram
pressure (DBP). However, this method can hinder the cir- (PPG) parameters to estimate blood pressure. They defined
culation of blood and is not intended for continuous BP PTT as the different time between the R wave of ECG and
measurement. the following pulse peak of PPG. They took the pulse wave
In recent years, great progress has been made in amplitude, PTT, pulse wave period, and area as the pa-
noninvasive continuous BP measurement. A study found rameters to send to a feed-forward back propagation neural
an inverse correlation between BP and the pulse transit network for training. Although the error could be within
time (PTT) [4, 5]. But this method requires an additional 5 mmHg, calibration was required before getting the output
[8]. Luo et al. fabricated a low-power (3 nW) piezoresistive (PS) in the radial artery. In PPG, light is emitted from a
sensor (PS) device and measured BP using a single pa- photodiode through a fingertip, a wrist, or the like, and the
rameter PTT but also required clinical calibration of the blood inside the artery absorbs the light and causes a change
initial values [9]. in the light intensity, so that a change in the amount of blood
With the advancement of sensor technology, piezoelectric can be detected, therefore the pulse wave is obtained. The
sensors nowadays become sensitive enough to reveal the pulse pulse wave signal in PPG is weak, and its SNR acquired by this
wave morphological features. However, during their use, method is generally not high. In addition, if a patient has
piezoelectric sensors require an external force to balance the arterial occlusion, tissue edema and blood clots, arrhythmia,
pressure between the probe and soft tissue [10]. In order to or weak peripheral circulation, this method will show con-
find the right amount of force applied to each individual, this siderable deviations. While in PS, a pressure sensor is applied
paper proposes a method to use body mass index (BMI) to to the skin above the radial artery and records the pulse
quantify the balance and find a suitable pressure for everyone. directly. The signal often has better SNR. Because one is an
This will be explained in the third part of this paper. indirect measure and the other is a direct measure, the
Our goal is to develop a stable, efficient continuous blood waveform characteristics collected by the two are different,
pressure measurement system that requires no user cali- and we have found that PS offers a lot of more details, thus
bration, to be used by patients with hypertension, cardio- more information, in its pulse waveform than in PPGs.
vascular, or other diseases for continuous BP monitoring. Figures 3(a) and 3(b) are pulse waves from a 25-year-old
The paper is organized as follows: in Section 2, the back- with arrhythmia, collected by PPG and PS, respectively. In
ground on propagation of pulse wave is given. In Section 3, a Figure 3(a), all the peaks except the main peak are all
two-domain fusion model and its use with artificial neural submerged due to the irregular heart beating. In compari-
networks (ANN) are elaborated. Section 4 shows experi- son, there are three peaks in addition to the main peak
mental results, and Section 5 provides a discussion and clearly visible in Figure 3(b). Figures 3(c) and 3(d) are the
conclusion on this proposed method. normal PPG and PS waveform from a 26-year-old female. It
can be seen that although the PPG can detect the pulse wave
2. Background (note the 2nd peak is submerged), the ratio between the 3rd
peak and the main peak is bigger than expected. The PS on
Figure 1 shows pulse wave and blood propagation. When the the other hand has a better representation for the true pulse
cardiac ejection is completed, the central aorta forms a wave wave. Therefore, we will use PS as the method of pulse wave
(#1), which is rebounded and generates a rebound wave (#2) measurement for our study.
when it propagates to the first reflection site (the arterial
node between thoracic aorta and renal arteries). The re-
bound wave will be rebounded again, producing a wave (#3) 3.2. Optimal Pressure for Piezoresistive Sensor. As for the
when the wave (#1) gets to the second reflection site (the pressure sensor for blood waveform measurement, we
arterial node between abdominal aorta and iliac arteries). choose Honeywell 1865 Series (Honeywell, Fort Mill, USA).
Finally, both of the rebound waves (#2 and #3) will su- It is a piezoresistive sensor that employs a solid-state pie-
perimpose on the main wave with a certain time delay and zopressure transducer mounted in a plastic package and
then propagate along the blood vessel to the radial artery or offers high resolution using its Wheatstone bridge strain
fingertip artery [11]. gauge design. It is designed to accurately collect tiny bi-
As shown in Figure 2, the features of the pulse wave ological signals with good linearity, high test accuracy, and
mainly include (b)–(g). The (b)–(e) points correspond to the fast response. It is used here to accurately detect the pressure
pulse wave’s starting point, main peak, trough point, and the change caused by pulsation.
peak produced by the superposition between the first re- The sensor is applied on top of the radial artery with a
bound wave (#2) and the main wave (#1). Points f and g are certain external pressure to obtain the pulse wave by
formed by the superposition between the second rebound detecting the beat of the superficial artery on the surface of
wave (#3) and the main wave (#1). Note that in a young the skin. In actual measurements, the pressure applied to a
person with good blood vessel elasticity, the rebound wave same person can be different at different times, resulting in a
can be weak and it can be submerged in the main wave and difference in waveforms for the same individual and possibly
not visible in the waveform. errors in BP estimation. We thus need to find a constant and
In order to accurately identify the characteristic points of optimal pressure applied onto the sensor for each individual.
the pulse wave, the pulse wave needs to be acquired with Unfortunately, the research on this problem is still relatively
high signal-to-noise ratio (SNR) and without distortion. The muted. We will thus address this issue first before we move
detection of pulse wave feature points is shown below in the onto our two-domain fusion model.
next section. Since the external force applied to the sensor is mainly to
counter the pressure from the elasticity of the soft tissue, we
3. Methods hypothesize that the optimal pressure for an individual is
dependent on the individual’s physical measures such as
3.1. Pulse Wave Detection. There are two mainstream height or obesity. We have thus done a lot of experiments to
methods for pulse wave detection. One is photoplethysmo- show this indeed is the case. When a force is applied to the
gram (PPG), and the other is to apply a piezoelectric sensor detection unit and varied from small to large, the soft tissue
Radial artery
#1
Aortic arch
#2 #1
#2
Thoracic aorta
#3 Renal arteries Reflection site I
Abdominal aorta #1
#3
Reflection site II
Time Iliac arteries
Figure 1: Relationship between pulse and BP.
0.176 Step 2. Place a piezoelectric film (Tekscan A201, Boston,

(c) USA) between the probe and the skin, adjust the wrist band
0.174
so the average pressure reading is 2 N, and store 20 sets of
(e) pressure value in 500 ms cycles. Then, withdraw the pie-
(d) zoelectric film and measure the pulse wave at this time.
0.172
Amplitude (V)
(g)
0.17 (f)
Step 3. Reinsert the piezoelectric film at the end of each
measurement. Increase the pressure by 0.5 N and repeat step
2. Stop when the pressure reaches 14 N.
0.168
(b′)
(b) Step 4. Find a set of pulse waves with ideal amplitude and
0.166
features and use median filter to fit the baseline of the
corresponding pressure measured by the film.
0.164
0 TC 20 40 60 80 100 120
TG Sample points Step 5. Find the value with the largest difference between
T
the pressure value and the baseline and record the value of
Figure 2: Standard pulse wave and characteristic points. the baseline at this time.
under the sensor probe forms a force field. If the pressure of Step 6. Repeat the above steps for 30 people and then use
the sensor probe and the arterial wall of the radial artery are the BMI and the baseline values obtained in step 5 to
in close equilibrium, the shape and features of the pulse wave perform a 3rd order polynomial fit to get the OP curve,
are the most conspicuous. Correspondingly, we define the where BMI is determined by the following equation:
pressure that produces the most detail of the pulse wave as
the optimal pressure (OP) for capturing pulse wave. Ad- weight(Kg)
BMI � . (1)
mittedly, if you want to get different OP for different people, height(m)2
you need to rely on some indicator to quantify the soft tissue
thickness above the radial artery. For this reason, we select In Figure 5(b), when the BMI is less than 24, the OP
the body mass index (BMI). curve is basically linear. It exhibits a saturation trend when
Figure 4(a) shows the piezoresistive sensor used in this the BMI goes beyond 24. Admittedly, the BMI of more than
article. It is equipped with a homemade casing and probe 24 is considered overweight in China [12]. So, we can come
(the contact point or the tip) and signal processing board. to a conclusion: when the BMI is lower than 24, the OP curve
Figures 4(b)–4(d) are pulse waves at the pressures 2.3 N, is linear with BMI. The OP is then determined by the fol-
3.8 N, and 14 N, respectively. At 2.3 N, the amplitude of the lowing equation:
pulse wave is small, and the features are not clear. On the OP � −0.0114 ∗ BMI3 + 0.7302 ∗ BMI2 − 15.0889 ∗ BMI
contrary, the pulse and its features are obvious at 3.8 N. Then
+ 104.4144.
at 14 N, the pulse wave is distorted with severe baseline shift
and waveform overlapping. In order to get the OP, we (2)
collected pulse waves for each person at different external
pressures. The steps are as follows:
3.3. Two-Domain Fusion Model. After the pulse wave is
acquired, we now move to the two-domain model. The
Step 1. Find the place where the radial artery beats the most. analysis and decomposition of pulse waves so far in the
Place the sensor probe on it. literature were essentially in a single domain, that is, either in
1.7 0.14
0.135
0.13
1.65
0.125
Amplitude (V)
Amplitude
0.12
0.115
1.6
0.11
0.105
1.55 0.1
40 60 80 100 120 140 160 180 200 100 150 200 250 300 350 400 450 500
Sample points Sample points
(a) (b)
2.45 0.18
0.178
2.4 0.176
0.174
Amplitude (V)
2.35 0.172
Amplitude
0.17
2.3 0.168
0.166
2.25 0.164
0.162
2.2 0.16
50 100 150 200 250 300 50 100 150 200 250 300 350 400 450 500
(c) (d)
Figure 3: Pulse waves captured by PS and PPG. (a) PPG with arrhythmia. (b) PS with arrhythmia. (c) Normal PPG. (d) Normal PS.
0.02
0.018
0.016
Plezoresistive 0.014
Amplitude (V)
sensor
0.012
0.01
0.008
0.006
0.004
0.002
200 400 600 800 1000 1200 1400
Sample points
(a) (b)
0.17 0.1
0.165 0.098
0.16 0.096
0.155 0.094
Amplitude (V)
Amplitude (V)
0.15 0.092
0.145 0.09
0.14 0.088
0.135 0.086
0.13 0.084
0.125 0.082
0.12 0.08
200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400
(c) (d)
Figure 4: Pulse wave with different pressure. (a) Piezoresistive sensor. (b) Pulse wave at 2.3 N. (c) Pulse wave at 3.8 N. (d) Pulse wave at 14 N.
5.6 6.5
5.4 6
Optimal pressure (N)
5.2 5.5
Amplitude (N)
5 5
4.8 4.5
4.6 4
4.4 3.5
0 2 4 6 8 10 12 14 16 19 20 21 22 23 24 25 26
Time (s) Value for BMI
Baseline
Pressure
(a) (b)
Figure 5: The pressure baseline and OP vs. BMI curve. (a) Baseline of pressure fitted with median filter. (b) Third-order polynomial fitting
for optimal pressure of capturing pulse wave.
the time domain or in the frequency domain. Millasseau In the time domain, each morphological point of the
et al. have obtained a generalized transfer function (GTF) by pulse wave can be related to physiological features that are in
the fast Fourier transform to analyze PPG and peripheral turn related to the elasticity of the arteries and the ability of
pressure pulse. They found that GTF is not only suitable for the heart to contract. One can derive blood pressure based
normotensive and hypertensive subjects, even adding few on these characteristics. In Figure 6, the wave will be
nitroglycerin (NTG, 500 mg sublingually) to the blood [13]. decomposed into 7 features in the time domain.
However, they only studied the peripheral pressure pulse In order to prevent confusion in the identification of
analysis but nothing for blood pressure estimation. Xing periodic features of several pulse waves, we extract features
et al. have decomposed the pulse wave in the frequency only in a single pulsation period and then identify other
domain and found that the blood pressure has a high features in other pulsation period the same way.
correlation with the amplitude of the fast Fourier transform At each systole, the amount of cardiac ejection is dif-
and that the phase also has a good correlation in the low ferent. The blood is transmitted through the arterial wave
frequency part [14]. and ultimately reflected in the amplitude of the systolic peak
0.18 0.18
0.178 (c)
0.176 0.175
0.174
Amplitude (V)
Amplitude (V)
0.172 0.17
(d)
0.17
(g)
0.168 0.165 (e)
S1 (f)
H
0.166
h
0.164 0.16
S2
0.162 (b′)
(b)
0.16 0.155
0 50 100 150 200 250 300 350 400 450 500 0 10 20 30 40 50 60 70 80 90 100
T1 Tf
Sample points T2
T
Figure 6: Feature points in the pulse wave in the time domain. Sample points
Figure 7: Illustration of time features.
[15]. Peripheral resistance will change with the diameter of
the artery and the viscosity of the blood, affecting the pulse
Table 1: Features extracted in time domain.
wave [16]. If the elasticity of the distal end of the blood vessel
is higher, the peak of the rebounding wave will be relatively Parameter Definition
lower, and the position superimposed on the main peak will FC The ratio between h and H
change the ratio of the area between the systolic and diastolic Sss The ratio between S1 and S1 plus S2
points, which in turn affects the blood pressure [17, 18]. Sds The ratio between S2 and S1 plus S2
Furthermore, vascular elasticity has an influence on the PTT. Tft The ratio between Tf and T
The ratio of time between the main peak and other peaks can Tst The ratio between T1 and T
be different if arteriosclerosis, tissue edema, and clots are SL The max slope between b and c
found out in the body [19]. The time domain focuses on
features from a single beat, such as the ratio of each peak dual-domain fusion model. Experimental details are given
within one pulse wave, while the beat-to-beat variations will below.
be taken into consideration in the frequency domain. Fig-
ure 7 and Table 1 show the area of S1 (cardiac systole) and S2
(cardiac diastole) and other relevant parameters in the time 3.4. Implementation. After fitting and obtaining OP for each
domain. people according to their BMI, a group of subjects un-
In the frequency domain, as shown in Figure 8, there are derwent measurement of pulse wave in the morning, af-
substantial amplitude components at 0.3 Hz, 1.2 Hz, 2.4 Hz, ternoon, and evening for 3 days. In addition, each
4.8 Hz, etc., after the fast Fourier transform (FFT) of the measurement is accompanied by BP obtained by a com-
pulse wave, but the amplitude decays rapidly even almost mercial medical upper-arm sphygmomanometer (Yuwell
submerge after 10 Hz. In this case, 0.3 Hz should be the 670B, Jiangsu, China). Both the features of pulse wave and
frequency of breathing, 1.2 Hz is the heart rate, and the BP are fed into a neural network for supervised learning. The
others are the multiple of the heart rate. It can be concluded process flow is shown in Figure 9 and is described below.
that the main components of the pulse wave in the frequency
domain are superimposed by the multiples or harmonics of Step 1. Collect real-time pulse wave data and then shape
the frequency of the heart rate and the frequency near the and filter to pulse wave.
heart rate. The amplitude component corresponding to the
frequency in the frequency domain of the pulse wave cor- Step 2. Decompose the pulse wave in the time and fre-
responds to the energy scale of each frequency. In order to quency domains.
improve accuracy and stability, this paper performs the FFT
on five complete pulsation cycles for better frequency
Step 3. Reconstruct and fuse features in the time and
sampling. The features in the frequency domain are the
frequency domains.
amplitudes and corresponding phases of the first three
characteristic peaks and are listed in Table 2. As for the other
points, their amplitudes are smaller and may be interfered Step 4. Perform the mean shift and normalization with the
with the pulse wave noise. recombined features.
The features from both domains are concatenated to-
gether and processed, combining the high precision of time Step 5. Send the processed features to the neural network to
domain and the stability of frequency domain into a better obtain SBP and DBP.
1.5 2
A1
1.5
1
1
0.5
Amplitude
Phase
0
A2
A3
–0.5
0.5
–1
P3
–1.5 P1
P2
0 –2
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Frequency Frequency
(a) (b)
Figure 8: Illustration of frequency features. (a) FFT amplitude of pulse wave. (b) FFT phase of pulse wave.
Table 2: Features extracted in frequency domain. shown in Figure 10, where x1 to xn are the spatial and
frequency features listed in Tables 1 and 2.
Parameter Definition In addition, the number of hidden layer nodes of the
A1 The amplitude of first peak neural network is quantitatively analyzed to study the ef-
A2 The amplitude of second peak fectiveness after the number of hidden layer neurons has
A3 The amplitude of third peak been increased. The results show that there is very little
P1 The phase of first peak
increase in accuracy when there are more than 30 hidden
P2 The phase of second peak
P3 The phase of third peak
layer nodes. Furthermore, for the accuracy to increase by
0.01 mmHg, the training time goes up an order of magnitude
by a network with a hidden layer of 50 neurons. Thus, we
Note in the two-domain fusion model, the fusion is done have chosen to use the aforementioned parameters for the
in the input level of the neural network. We first exact ANN.
features from the time and the frequency domains separately
to form two vectors and then combine them into one data
matrix to feed into the neural network for training. An 4. Experiment Results
additional label matrix records the true blood pressure
measured by the commercial sphygmomanometer to serve A total of 30 volunteers have participated in the experiment,
as the training target. The neural network then iterates and a total of 1231 pulse wave data were measured. In order
through the combined feature data, from which the weight to improve the accuracy and stability of blood pressure
corresponding to each feature is generated. The final trained estimation, we collected the pulse waves in a single cycle.
model will contain coefficients for all time and frequency Besides, we built a threshold matrix to prevent any pa-
features, thus achieving the two-domain fusion. rameters, such as period and amplitude, exceeding the
In this paper, the quasi-Newton method is adopted for threshold range, upon which the data are considered ab-
better convergence speed of the training. The neural network normal and discarded. The training and fitting results of BP
is a feed-forward back propagation ANN with 12 input are shown in Figure 11.
points, 30 nodes in single hidden layer, and 2 output nodes. The training automatically stopped after the about 200
All 12 inputs of the network are the features extracted from iterations when the verification error became minimal. The
the time domain and the frequency domain. The major total errors were within ±2 mmHg according to the error
superiorities of this type of network are as follows: (1) it is histogram (Figure 11(d)). In order to verify the performance
robust. Error in one input does little damage to the other of the model, we used the 15% data previously put aside and
inputs. (2) It can be proved mathematically that a three-layer tested it. It is found that the output and target values show a
network can approximate any continuous nonlinear re- highly linear correlation.
lationship with arbitrary precision. (3) It has a strong self- In addition, we also compared our model results with
adaptive learning ability and outputs adaptive learning that of the traditional single-domain model. The perfor-
contents to the weights of networks [8, 20]. In order to verify mance of the single-domain model is inferior to that of our
the training results, the data are randomly divided into 70%, two-domain model. As shown in Figure 12, the error is
15%, and 15%, used for training, verification, and testing of ±3 mmHg and ±9 mmHg in the time domain and the fre-
the neural network, respectively. The structure of ANN is quency domain, respectively.
Time
domain
decomposit
-ion SBP
Pulse Feed-
wave data Shaping Two-domain forward
and parameter Preprocessing back
filtering fusion propagation
ANN DBP
Frequency
domain
decomposit
-ion
Figure 9: System structure.
x1
x2
SBP
x3
DBP
xn
Figure 10: Structure of ANN.
Training: R = 1 Test: R = 0.99996
120 120
Output = 1 ∗ target ± 0.00071
Output = 1 ∗ target ± 0.00092
110 110
100 100
90 90
80 80
70 70
60 60
60 80 100 120 60 80 100 120
Target Target
Data Data
Fit Fit
Y=T Y=T
(a) (b)
Best validation performance is 0.0093008 at epoch 206 Error histogram with 20 bins
350
102
Mean squared error (mse)
300
250
Instances
100 200
150
100
10–2
50
–0.8734
–0.8059
–0.7385
–0.671
–0.6036
–0.5361
–0.4687
–0.4012
–0.3338
–0.2663
–0.1989
–0.1314
–0.06398
0.003466
0.07091
0.1384
0.2058
0.2733
0.3407
0.4082
0 20 40 60 80 100 120 140 160 180 200
212 epochs
Train Test
Validation Best Errors = targets – outputs
Training Test
Validation Zero error
(c) (d)
Figure 11: Two-domain model result. (a) Training error. (b) Testing error. (c) Iterations of ANN. (d) Two-domain error.
Error histogram with 20 bins Error histogram with 20 bins
20 25
20
15
Instances
Instances
15
10
10
5
5
0 0
–7.953
–7.061
–6.17
–5.278
–4.386
–3.494
–2.602
–1.711
–0.8189
0.07288
0.9647
1.856
2.748
3.64
4.532
5.424
6.315
7.207
8.099
8.991
–2.04
–1.792
–1.545
–1.297
–1.05
–0.8023
–0.5548
–0.3072
–0.05973
0.1878
0.4353
0.6828
0.9303
1.178
1.425
1.673
1.92
2.168
2.415
2.663
Errors = targets – outputs Errors = targets – outputs

Training Test Training Test
Validation Zero error Validation Zero error
(a) (b)
Figure 12: Errors of single-domain model. (a) Error of time domain. (b) Error of frequency.
5. Discussion and Conclusion model proposed in this paper not only retains the high
accuracy of the time domain model but also integrates the
Experiments show that piezoresistive sensors can very well stability of the frequency domain model. Experiments on the
detect the features of the pulse wave under complex factors time and frequency domains and two domains combined
such as arrhythmia and are superior to PPG. There exists an were done and compared. The results show that the con-
optimal restraint pressure for the piezoresistive sensor, and tinuous blood pressure estimation based on the two-domain
it can be related back to the BMI for different people. model can give BP with higher accuracy and stability than a
Considering the fact that the time domain model is single-domain model does.
unstable when measuring pulse wave data but has a high- There are limitations to this study, however. For ex-
accuracy contribution to BP and that the frequency domain ample, the training and testing sample size is still small and
model is stable in decomposing pulse waves, the two-domain not diverse enough. Subjects were not divided into training
and testing groups, i.e., data from all subjects were mixed eletrocardiogram and photoplethysmogram by Back-
together and then divided for training and testing, which propagation neural network,” Computers in Industry,
could introduce internal correlation. In addition, the fea- vol. 89, no. 5, pp. 50–59, 2017.
tures we choose to extract and use may not be the best to [9] N. Luo, W. Dai, C. Li et al., “Flexible piezoresistive sensor
represent the underlying mechanisms of BP, and the number patch enabling ultralow power cuffless blood pressure mea-
surement,” Advanced Functional Materials, vol. 26, no. 8,
of features may not be the best. In the future, we will
pp. 1178–1187, 2016.
continue to optimize the feature selection in both the time [10] J. Li, “Pulse wave analysis under different pressure,” Master’s
and frequency domains and plan to enlarge the sample space thesis, Yanshan University, Qinhuangdao, China, 2015.
and amend the method of data splitting for a more rigid [11] M. C. Baruch, D. E. R. Warburton, S. S. D. Bredin, A. Cote,
evaluation study. D. W. Gerdt, and C. M. Adkins, “Pulse decomposition
analysis of the digital arterial pulse during hemorrhage
Data Availability simulation,” Nonlinear Biomedical Physics, vol. 5, no. 1, p. 1,
2015.
The datasets generated and analyzed during the current [12] B. Zhou, “The obesity standard of Chinese people,” Popular
study are available from the corresponding author upon medicine, vol. 206, p. 7, 2002.
reasonable request, without breaching participant [13] S. C. Millasseau, F. G. Guigui, R. P. Kelly et al., “Noninvasive
assessment of the digital volume pulse,” Hypertension, vol. 36,
confidentiality.
no. 6, pp. 952–956, 2000.
[14] X. Xing and M. Sun, “Optical blood pressure estimation with
Conflicts of Interest photoplethysmography and FFT-based neural networks,”
Biomedical Optics Express, vol. 7, no. 8, pp. 3007–3020, 2016.
The authors declare that they have no conflicts of interest [15] Z. Yuan, X. Huang, and F. Fan, “Analysis of photo-
regarding the publication of this manuscript. plethysmogram on different positions of 253 normal adults,”
Journal Hunan College of Traditional Chinese Medicine,
Acknowledgments vol. 20, no. 3, pp. 1–3, 2000.
[16] M. Elgendi, “On the analysis of fingertip photoplethysmo-
This study was supported in part by Grant 11805263 from gram signals,” Current Cardiology Reviews, vol. 8, no. 1,
NSFC, Grant 2018K058B from Jiangsu Post-Doctoral Funds, pp. 14–25, 2012.
and the 5th 333 Project of Jiangsu Province of China. [17] M. Saito, M. Matsukawa, T. Asada, and Y. Watanabe,
“Noninvasive assessment of arterial stiffness by pulse wave
analysis,” IEEE Transactions on Ultrasonics, Ferroelectrics, and
References Frequency Control, vol. 59, no. 11, pp. 2411–2419, 2012.
[18] P. Shi, S. Hu, Y. Zhu, J. Zheng, Y. Qiu, and P. Y. S. Cheang,
[1] M. H. Forouzanfar, P. Liu, G. A. Roth et al., “Global burden of “Insight into the dicrotic notch in photoplethysmographic
hypertension and systolic blood pressure of at least 110 to 115 pulses from the finger tip of young adults,” Journal of Medical
mmHg,” JAMA, vol. 317, no. 2, pp. 165–182, 2017. Engineering & Technology, vol. 33, no. 8, pp. 628–633, 2009.
[2] S. Lewington, R. Clarke, N. Qizilbash, R. Peto, R. Collins, and [19] H. Lax, A. W. Feinberg, and B. M. Cohen, “Studies of the
Prospective Studies Collaboration, “Age-specific relevance of arterial pulse wave,” Journal of Chronic Diseases, vol. 3, no. 6,
usual blood pressure to vascular mortality: a meta-analysis of pp. 618–631, 1956.
individual data for one million adults in 61 prospective [20] S. C. Gao, P. Wittek, L. Zhao, and W. J. Jiang, “Data-driven
studies,” Lancet, vol. 360, no. 9349, pp. 1903–1913, 2002. estimation of blood pressure using the photoplethysmo-
[3] R. Agarwal, J. E. Bills, T. J. W. Hecht, and R. P. Light, “Role of graphic signals,” in Proceedings of 2016 38th Annual In-
home blood pressure monitoring in overcoming therapeutic ternational Conference of the IEEE Engineering in Medicine
inertia and improving hypertension control: a systematic and Biology Society (EMBC), Orlando, FL, USA, August 2016.
review and meta-analysis,” Hypertension, vol. 57, no. 1,
pp. 29–38, 2011.
[4] L. A. Geddes, M. H. Voelz, C. F. Babbs, J. D. Bourland, and
W. A. Tacker, “Pulse transit time as an indicatoro of arterial
blood pressure,” Psychophysiology, vol. 18, no. 1, pp. 71–74,
1981.
[5] Y. Choi, Q. Zhang, and S. Ko, “Noninvasive cuffless blood
pressure estimation using pulse transit time and Hilbert-
Huang transform,” Computers and Electrical Engineering,
vol. 39, no. 1, pp. 103–111, 2013.
[6] C. M. Wu, C. Y. Chuang, Y. J. Chen, and S. C. Chen, “A new
estimate technology of non-invasive continuous blood
pressure measurement based on electrocardiograph,” Ad-
vances in Mechanical Engineering, vol. 8, no. 6, article
168781401665368, 2016.
[7] M. F. O’Rourke, “Time domain analysis of the arterial pulse in
clinical medicine,” Medical and Biological Engineering and
Computing, vol. 47, no. 2, pp. 119–129, 2009.
[8] Z. Xu, J. Liu, X. Chen, Y. Wang, and Z. Zhao, “Continuous
blood pressure estimation based on multiple parameters from
Hindawi
https://doi.org/10.1155/2018/7325938
Research Article
Some Similarity Measures of Neutrosophic Sets Based on the
Euclidean Distance and Their Application in Medical Diagnosis
Donghai Liu ,1 Guangyan Liu,1 and Zaiming Liu2

1
Department of Mathematics, Hunan University of Science and Technology, Xiangtan, Hunan 411201, China
2
Department of Mathematics, Central South University, Changsha, Hunan 410075, China
Correspondence should be addressed to Donghai Liu; donghailiu@126.com
Received 6 July 2018; Revised 25 September 2018; Accepted 23 October 2018; Published 28 November 2018
Copyright © 2018 Donghai Liu et al. This is an open access article distributed under the Creative Commons Attribution License,
Similarity measure is an important tool in multiple criteria decision-making problems, which can be used to measure the
difference between the alternatives. In this paper, some new similarity measures of single-valued neutrosophic sets (SVNSs) and
interval-valued neutrosophic sets (IVNSs) are defined based on the Euclidean distance measure, respectively, and the proposed
similarity measures satisfy the axiom of the similarity measure. Furthermore, we apply the proposed similarity measures to
medical diagnosis decision problem; the numerical example is used to illustrate the feasibility and effectiveness of the proposed
similarity measures of SVNSs and IVNSs, which are then compared to other existing similarity measures.
1. Introduction and the indeterminacy-membership degree πE (xi ) � 1 −

uE (xi ) − vE (xi ). The IFS is more effective to deal with the
The concept of fuzzy set (FS) A � 􏼈〈xi , uA (xi )〉|xi ∈ X􏼉 in vague information than the FS and IVFS. Then, the in-
X � 􏼈x1 , x2 , . . . , xn 􏼉 was proposed by Zadeh [1], where the formation about the admission of the student can be rep-
membership degree uA (xi ) is a single value between zero resented as an IFS E � 0.5, 0.3, 0.2, where 0.5, 0.3, and 0.2
and one. The FS has been widely applied in many fields, stand for the membership degree, nonmembership degree,
such as medical diagnosis, image processing, supply de- and indeterminacy-membership degree, respectively.
cision-making [2–4], and so on. In some uncertain decision- However, the IFS also have limitation in expressing the
making problems, the degree of membership is assumed not decision information. For example, three groups of experts
exactly as a numerical value but as an interval. Hence, Zadeh evaluate the benefits of the stock, a group of experts thinks
[5] proposed the interval-valued fuzzy set (IVFS). However, the possibility of the stock that will be profitable is 0.6, the
the FS and IVFS only have the membership degree, and they second group of experts thinks the possibility of loss is 0.3,
cannot describe the nonmembership degree of the element the third group of experts is not sure whether the stock that
belonging to the set. For example, in the national entrance will be profitable is 0.4. In this case, the IFS cannot express
examination for postgraduate, a panel of ten professors such information because 0.6 + 0.3 + 0.4 > 1. Therefore,
evaluated the admission of a student; five professors con- Wang et al. [7] proposed a single-valued neutrosophic set
sidered the student can be accepted, three professors dis- (SVNS) N � 􏼈〈xi , TN (xi ), IN (xi ), FN (xi )〉|xi ∈ X􏼉, where
approved of his or her admission, and two professors TN (xi ), IN (xi ) and FN (xi ) represent the degree of the
remained neutral. In this case, the FS and IVFS cannot truth-membership, indeterminacy-membership, and falsity-
represent such information. In order to solve this problem, membership, respectively, and they belong to [0, 1]. So, the
Atanassov et al. [6] proposed the intuitionistic fuzzy set information about the benefits of the stock can be repre-
(IFS) E � 􏼈〈xi , uE (xi ), vE (xi )〉|xi ∈ X􏼉, where uE (xi )(0 ≤ sented as N � 0.6, 0.4, 0.3. However, due to the uncertainty
uE (xi ) ≤ 1) and vE (xi )(0 ≤ vE (xi ) ≤ 1) represent the mem- of the decision-making environment in multiple criteria
bership degree and nonmembership degree, respectively, decision-making problems, the single numerical value
cannot meet the needs of evaluating information. Then, 2.1. SVNS

Wang [8] defined the interval-valued neutrosophic set
(IVNS) based on the SVNS, which used the interval to Definition 1. Given a fixed set X � 􏼈x1 , x2 , . . . , xn 􏼉 [7], the
describe truth membership degree, indeterminacy mem- SVNS N in X is defined as follows:
bership degree, and falsity membership degree, respectively. 􏼌􏼌
N � 􏽮􏼊xi , TN xi 􏼁, IN xi 􏼁, FN xi 􏼁􏼋􏼌􏼌xi ∈ X􏽯, (1)
Since the neutrosophic set was proposed, there have been
some researchers focusing on this subject [9–12]. where the function TN (xi ) : X ⟶ [0, 1] defines the truth-
On the other hand, similarity measure is an important membership degree, the function IN (xi ) : X ⟶ [0, 1]
tool in multiple criteria decision-making problems, which defines indeterminacy-membership degree, and the func-
can be used to measure the difference between the alter- tion FN (xi ) : X ⟶ [0, 1] defines the falsity-membership
natives. Many studies about the similarity measure are degree, respectively. For any SVNS N, it holds that
obtained. For example, Beg et al. [13] proposed a similarity 0 ≤ TN (xi ) + IN (xi ) + FN (xi ) ≤ 3 (∀xi ∈ X).
measure of FSs based on the concept of ϵ − fuzzy transitivity For any two SVNSs N1 � {〈xi , TN1 (xi ), IN1 (xi ), FN1
and discussed the degree of transitivity of different similarity (xi )〉|xi ∈ X} and N2 � {〈xi , TN2 (xi ), IN2 (xi ), FN2 (xi )〉
measures. Song et al. [14] considered the similarity measure |xi ∈ X}}, the following properties are satisfied:
of IFSs and proposed corresponding distance measure be-
tween intuitionistic fuzzy belief functions. Majumdar and (1) N1 ⊆ N2 if and only if TN1 (xi ) ≤ TN2 (xi ), IN1
Samanta [15] proposed a similarity measure between SVNSs (xi ) ≥ IN2 (xi ), and FN1 (xi ) ≥ FN2 (xi )
based on the membership degree. (2) N1 � N2 if and only if N1 ⊆ N2 and N2 ⊆ N1
In addition, cosine similarity measure is also an im-
portant similarity measure, and it can be defined as the
inner product of two vectors divided by the product of their
lengths. There are some scholars who study the cosine 2.2. IVNS
similarity measures [16–21]. For example, Ye [16] proposed
the cosine similarity measure and weighted cosine simi- Definition 2. Given a fixed set X � 􏼈x1 , x2 , . . . , xn 􏼉 [8], the
larity measure of IVFSs with risk preference, and they were IVNS N′ on X is defined as follows:
applied to the supplier selection problem. Then, Ye [17] N′ � 􏼚􏼜xi , 􏽨TLN′ xi 􏼁, TU L U
N′ xi 􏼁􏽩, 􏽨IN′ xi 􏼁, IN′ xi 􏼁􏽩,
proposed the cosine similarity measure of IFSs and (2)
applied it to medical diagnosis and pattern recognition. L U
􏼌􏼌
􏽨FN′ xi 􏼁, FN′ xi 􏼁􏽩􏼝􏼌􏼌xi ∈ X􏼛,
Furthermore, Ye [18] defined the cosine similarity measure
of SVNSs and IVNSs, but when the SVNSs N1 ≠ N2 ,
where TN′ (xi ) � [TLN′ (xi ), TU (xi )], IN′ (xi ) � [ILN′ (xi ),
cos(N1 , N2 ) � 1 (the example can be seen in Section 3). N′
IN′ (xi )], and FN′ (xi ) � [FN′ (xi ), FU
U L
(xi )] represent the
Furthermore, Ye [19] proposed the improved cosine N′
truth-membership function, the indeterminacy-member-
similarity measures of SVNSs and IVNSs based on cosine
ship function, and the falsity-membership function, re-
function.
spectively. For any xi ∈ X, it holds that TN′ (xi ), IN′ (xi ),
In this paper, we propose a new method to construct the
FN′ (xi ) ⊆ [0, 1] and 0 ≤ TU (xi ) + IU (xi ) + FU (xi ) ≤ 3.
similarity measures of SVNSs, which is based on the existing N′ N′ N′
For any two IVNSs N1′ � {〈xi , [TLN′ (xi ), TU (xi )],
similarity measure proposed by Majumdar and Samanta [15] 1 N′1
L U L U
and Ye [18], respectively. They play an important role in [IN′ (xi ), IN′ (xi )], [FN′ (xi ), FN′ (xi )]〉|xi ∈ X} and N2′
1 1 1 1
practical application, especially in pattern recognition, � {〈xi , [TLN′ (xi ), TU N′ (xi )], [ILN′ (xi ), IU
N′2
(xi )], [FLN′ (xi ), FUN′2
2 2 2 2
medical diagnosis, and so on. Furthermore, we will propose (xi )]〉 |xi ∈ X}, the following properties are satisfied:
the corresponding similarity measures of IVNSs.
The rest of the paper is organized as follows. In Section (1) N1′ ⊆N2′ if and only if TLN′ (xi )≤TN′2 L (xi ),TU N′1
(xi )≤
1
U L L U U L
2, the basic definition and some properties about SVNS and TN′ (xi ),IN′ (xi )≥IN′ (xi ),IN′ (xi )≥IN′ (xi ),FN′ (xi )≥
2 1 2 1 2 1
IVNS are given. In Section 3, we proposed a method FLN′ (xi ),and FU (xi )≥FU (xi )
N ′ N ′
to construct the new similarity measures of SVNSs and 2 1 2
(2) N1′ � N2′ if and only if N1′ ⊆ N2′ and N2′ ⊆ N1′
IVNSs, respectively. In Section 4, we apply the proposed
new similarity measures to medical diagnosis problems, the
numerical examples are used to illustrate the feasibility and Remark 1. When TLN′ (xi ) � TU N′1
(xi ), ILN′ (xi ) � IU
N′1
(xi ), FLN′
1 1 1
effectiveness of the proposed similarity measures, which are (xi ) � FU (xi ), the IVNS N1′ is reduced to the SVNS N1 .
N′
then compared to other existing similarity measures. Fi- 1
nally, the conclusions and future studies are discussed in

Section 5. 2.3. Existing Distance Measures between SVNSs and IVNSs
Definition 3. Let N1 � {〈xi , TN1 (xi ), IN1 (xi ), FN1 (xi )〉

2. Preliminaries |xi ∈ X} and N2 � {〈xi , TN2 (xi ), IN2 (xi ), FN2 (xi )〉|xi ∈X}}
In this section, we give some basic knowledge about the be any two SVNSs in X � {x1 , x2 , . . . , xn } [15]; then, the
SVNS and the IVNS. Some existing distance measures are Euclidean distance between SVNSs N1 and N2 is defined as
also introduced, which will be used in the next section. follows:
􏽶��
􏽴
2 2 2
􏽐ni�1 􏼔􏼐TN1 xi 􏼁 − TN2 xi 􏼁􏼑 + 􏼐IN1 xi 􏼁 − IN2 xi 􏼁􏼑 + 􏼐FN1 xi 􏼁 − FN2 xi 􏼁􏼑 􏼕 (3)
DSVNS N1 , N2 􏼁 � .
3n
Definition 4. Let N1′ � {〈xi , [TLN′ (xi ), TU N′1

(xi )], [ILN′ (xi ), be any two IVNSs in X � 􏼈x1 , x2 , . . . , xn 􏼉 [22]; the Euclidean
1 1
IU
N′1
(xi )], [FLN′ (xi ), FU
N′1
(xi )]〉|xi ∈ X} and N2′ � {〈xi , [TLN′ distance between IVNSs N1′ and N2′ is defined as follows:
1 2
(xi ), TU
N′
(xi )], [ILN′ (xi ), IU
N′
(xi )], [FLN′ (xi ), FU
N′
(xi )]〉|xi ∈ X}
2 2 2 2 2
DIVNS N1′, N2′􏼁

􏽶��
􏽴 2 2 2 2 2 2
􏽐ni�1 􏼢􏼒TLN′ xi 􏼁 − TLN′ xi 􏼁􏼓 + 􏼒TU
N′
xi 􏼁 − TU
N′
xi 􏼁􏼓 + 􏼒ILN′ xi 􏼁 − ILN′ xi 􏼁􏼓 + 􏼒IU
N′
xi 􏼁 − IU
N′
xi 􏼁􏼓 + 􏼒FLN′ xi 􏼁 − FLN′ xi 􏼁􏼓 + 􏼒FU
N′
xi 􏼁 − FU
N′
xi 􏼁􏼓 􏼣
1 2 1 2 1 2 1 2 1 2 1 2
� .
6n
(4)
Next, we propose a new method to construct the sim- (3) S(N1 , N2 ) � S(N2 , N1 ).
ilarity measures of SVNSs and IVNSs based on the Euclidean
Then, the similarity measure S(N1 , N2 ) is a genuine
distance measure.
similarity measure.
3. Several New Similarity Measures

3.1. The New Similarity Measures between SVNSs. To in-
The similarity measure is a most widely used tool to evaluate
troduce the new similarity measure between SVNSs, we
the relationship between two sets. The following axiom
first review the similarity measure S1SVNS between N1 and
about the similarity measure of SVNSs (or IVNSs) should be
N2 defined by Majumdar et al. [15], which is given as
satisfied:
follows:
Lemma 1. Let X � 􏼈x1 , x2 , . . . , xn 􏼉 be the universal set [18]
if the similarity measure S(N1 , N2 ) between SVNSs (or Definition 5. Let X � 􏼈x1 , x2 , . . . , xn 􏼉 be a universal set [15],
IVNSs) N1 and N2 satisfies the following properties: for any two SVNSs N1 � {〈xi , TN1 (xi ), IN1 (xi ), FN1 (xi )〉
|xi ∈ X} and N2 � {〈xi , TN2 (xi ), IN2 (xi ), FN2 (xi )〉 |xi ∈ X};
(1) 0 ≤ S(N1 , N2 ) ≤ 1 the similarity measure of SVNSs between N1 and N2 is
(2) S(N1 , N2 ) � 1 if and only if N1 � N2 defined as follows:
􏽐ni�1 􏼐min 􏼐TN1 xi 􏼁, TN2 xi 􏼁􏼑 + min 􏼐IN1 xi 􏼁, IN2 xi 􏼁􏼑 + min 􏼐FN1 xi 􏼁, FN2 xi 􏼁􏼑􏼑
S1SVNS N1 , N2 􏼁 � . (5)
􏽐ni�1 􏼐max 􏼐TN1 xi 􏼁, TN2 xi 􏼁􏼑 + max 􏼐IN1 xi 􏼁, IN2 xi 􏼁􏼑 + max 􏼐FN1 xi 􏼁, FN2 xi 􏼁􏼑􏼑
It is already known that the similarity measure S1SVNS 1

defined by Majumdar et al. [15] satisfies the properties S∗1SVNS N1 , N2 􏼁 � S N , N 􏼁 + 1 − DSVNS N1 , N2 􏼁􏼁.
2 1SVNS 1 2
in Lemma 1. It is proposed based on the membership (6)
degree; in this section, we adopt the various methods for
calculating the similarity measure between neutrosophic The proposed similarity measure of SVNSs satisfies the
sets. following Theorem 1.
Firstly, we propose a new method to construct a new
similarity measure of SVNSs, which is based on the simi- Theorem 1. The similarity measure S∗1SVNS (N1 , N2 ) between
larity measure proposed by Majumdar et al. [15] and the N1 � {〈xi , TN1 (xi ), IN1 (xi ), FN1 (xi )〉|xi ∈ X} and N2 �
Euclidean distance; it can be defined as follows: {〈xi , TN2 (xi ), IN2 (xi ), FN2 (xi )〉|xi ∈ X} satisfies the follow-
ing properties:
Definition 6. Let X � 􏼈x1 , x2 , . . . , xn 􏼉 be a universal set, for
(1) 0 ≤ S∗1SVNS (N1 , N2 ) ≤ 1
any two SVNSs N1 � {xi , TN1 (xi ), IN1 (xi ), FN1 (xi )|xi ∈ X}
and N2 � {xi , TN2 (xi ), IN2 (xi ), FN2 (xi )|xi ∈ X}; a new sim- (2) S∗1SVNS (N1 , N2 ) � 1 if and only if N1 � N2
ilarity measure S∗1SVNS (N1 , N2 ) is defined as follows: (3) S∗1SVNS (N1 , N2 ) � S∗1SVNS (N2 , N1 )
Proof. On the other hand, when N1 � N2 , according to for-

mulae (3) and (5) DSVNS (N1 , N2 ) � 0 and S1SVNS (N1 , N2 ) �
(1) Because DSVNS (N1 , N2 ) is an Euclidean distance 1 are obtained respectively. Furthermore, we can get
measure, obviously, 0 ≤ DSVNS (N1 , N2 ) ≤ 1. Fur- S∗1SVNS (N1 , N2 ) � 1.
thermore, according to Proposition 4.2.2 by
Majumdar et al. [15], we know 0 ≤ S1SVNS (N1 , (3) S∗1SVNS (N1 , N2 ) � S∗1SVNS (N2 , N1 ) is straightforward.
N2 ) ≤ 1. Then, 0 ≤ 1/2(S1SVNS (N1 , N2 ) + 1 − DSVNS From Theorem 1, we know the proposed new simi-
(N1 , N2 )) ≤ 1, i.e., 0 ≤ S∗1SVNS (N1 , N2 ) ≤ 1. larity measure S∗1SVNS (N1 , N2 ) is a genuine similarity
(2) If S∗1SVNS (N1 , N2 ) � 1, we have S1SVNS (N1 , N2 )+ measure.
1 − DSVNS (N1 , N2 ) � 2, that is, S1SVNS (N1 , N2 ) � On the other hand, cosine similarity measure is also
1 + DSVNS (N1 , N2 ). Because DSVNS (N1 , N2 ) is the an important similarity measure. In 2014, Ye [18] pro-
Euclidean distance measure, 0 ≤ DSVNS (N1 , N2 ) ≤ 1. posed a cosine similarity measure between SVNSs as
Furthermore, 0 ≤ S1SVNS (N1 , N2 ) ≤ 1 is obtained in follows: □
Proposition 4.2.2 [15], then S1SVNS (N1 , N2 ) � 1 and
DSVNS (N1 , N2 ) � 0 should be established at the same
Definition 7. Let X � 􏼈x1 , x2 , . . . , xn 􏼉 be a universal set
time. If the Euclidean distance measure DSVNS
[18], for any two SVNSs N1 � {〈xi , TN1 (xi ), IN1 (xi ),
(N1 , N2 ) � 0, N1 � N2 is obvious. According to
FN1 (xi )〉|xi ∈ X} and N2 � {〈xi , TN2 (xi ), IN2 (xi ), FN2 (xi )〉
Proposition 4.2.2 by Majumdar et al. [15], when
|xi ∈ X}, the cosine similarity measure between N1 and N2 is
S1SVNS (N1 , N2 ) � 1, N1 � N2 ; so, if S∗1SVNS (N1 ,
defined as follows:
N2 ) � 1, N1 � N2 is obtained.
1 n TN1 xi 􏼁TN2 xi 􏼁 + IN1 xi 􏼁IN2 xi 􏼁 + FN1 xi 􏼁FN2 xi 􏼁

S2SVNS N1 , N2 􏼁 � 􏽘 􏽱�� 􏽱��. (7)
n i�1 T x 􏼁 + I2 x 􏼁 + F2 x 􏼁 T2 x 􏼁 + I2 x 􏼁 + F2 x 􏼁
2
N1 i N1 i N1 i N2 i N2 i N2 i
From Example 1, we know the cosine similarity measure 0.2, 0.1, 0.3, we have S∗2SVNS (N1 , N2 ) � 0.8920. We can see
defined by Ye [18] does not satisfy Lemma 1. that the proposed new similarity measure S∗2SVNS (N1 , N2 )
overcomes the shortcoming of cosine similarity measure
Example 1. For two SVNSs N1 � x, 0.4, 0.2, 0.6 and S2SVNS (N1 , N2 ) defined by Ye [18].
N2 � x, 0.2, 0.1, 0.3, we can easily know N1 ≠ N2 . But
using formula (7) to calculate the cosine similarity measure Theorem 2. The similarity measure S∗2SVNS (N1 , N2 ) between
S2SVNS (N1 , N2 ), we have S2SVNS (N1 , N2 ) � 1. That is to say, N1 � {〈xi , TN1 (xi ), IN1 (xi ), FN1 (xi )〉|xi ∈ X} and N2 � {〈xi ,
when N1 ≠ N2 , S2SVNS (N1 , N2 ) � 1, which means the cosine TN2 (xi ), IN2 (xi ), FN2 (xi )〉|xi ∈ X} satisfies the following
similarity measure S2SVNS (N1 , N2 ) defined by Ye [18] does properties:
not satisfy the necessary condition of property 2 in Lemma 1;
thus, it is not a genuine similarity measure. Furthermore, Ye (1) 0 ≤ S∗2SVNS (N1 , N2 ) ≤ 1
[19] proposed the improved cosine similarity measures of (2) S∗2SVNS (N1 , N2 ) � 1 if and only if N1 � N2
SVNS based on the cosine similarity measure proposed by (3) S∗2SVNS (N1 , N2 ) � S∗2SVNS (N2 , N1 )
Ye [18], which overcomes its shortcoming.
In this paper, we go on proposing another new similarity
measure of SVNSs based on the cosine similarity measure Proof. The proof of the properties (1) and (3) are similar to
proposed by Ye [18] and the Euclidean distance DSVNS . It Theorem 1; here, we only give the proof of property (2).
considers the similarity measure not only from the point of If S∗2SVNS (N1 , N2 ) � 1, we have S2SVNS (N1 , N2 ) + 1 −
view of algebra but also from the point of view of geometry, DSVNS (N1 , N2 ) � 2, i.e., S2SVNS (N1 , N2 ) � 1 + DSVNS (N1 ,
which can be defined as: N2 ). Because DSVNS (N1 , N2 ) is the Euclidean distance
measure, 0 ≤ DSVNS (N1 , N2 ) ≤ 1. According to the property
Definition 8. Let X � 􏼈x1 , x2 , . . . , xn 􏼉 be a universal set, for of S2SVNS (N1 , N2 ) in Ye [18], 0 ≤ S2SVNS (N1 , N2 ) ≤ 1; then,
any two SVNSs N1 � {〈xi , TN1 (xi ), IN1 (xi ), FN1 (xi )〉 S2SVNS (N1 , N2 ) � 1 and DSVNS (N1 , N2 ) � 0 should be held
|xi ∈ X} and N2 � {〈xi , TN2 (xi ), IN2 (xi ), FN2 (xi )〉|xi ∈ X}; at the same time. When S2SVNS (N1 , N2 ) � 1, we have
a new similarity measure S∗2SVNS (N1 , N2 ) is defined as follows: TN1 (xi ) � k · TN2 (xi ), IN1 (xi ) � k · IN2 (xi ), and FN1 (xi ) �
1 k · FN2 (xi ) (k is a constant). When DSVNS (N1 , N2 ) � 0, we
S∗2SVNS N1 , N2 􏼁 � S2SVNS N1 , N2 􏼁 + 1 − DSVNS N1 , N2 􏼁􏼁. have N1 � N2 . Then N1 � N2 is obtained.
2
(8) On the other hand, according to formulae (3) and (7), if
N1 � N2 , DSVNS (N1 , N2 ) � 0 and S2SVNS (N1 , N2 ) � 1 are
obtained, respectively; then we can get S∗2SVNS (N1 , N2 ) � 1.
Remark 2. Using formula (8) to calculate Example 1 Thus, S∗2SVNS (N1 , N2 ) satisfies all the properties in
again, for two SVNSs N1 � x, 0.4, 0.2, 0.6 and N2 � x, Theorem 2. □
3.2. Some New Similarity Measures between IVNSs. In some Definition 9. Let X � 􏼈x1 , x2 , . . . , xn 􏼉 be a universal set, for
situations, it is difficult to provide the truth-membership any two IVNSs N1′ � {〈xi , [TLN′ (xi ), TU N′1
(xi )], [ILN′ (xi ),
1 1
degree, false-membership degree, and indeterminate-mem- IN′ (xi )], [FN′ (xi ), FN′ (xi )]〉|xi ∈ X} and N2′ � {〈xi , [TLN′
U L U
1 1 1 2
bership degree with a precise numerical value; Wang [8] used (xi )TU (xi )], [ILN′ (xi ), IU (xi )], [FLN′ (xi ), FU (xi )]〉 |xi ∈ X}
N′2 2 N′2 2 N′2
the interval numbers to express the related membership [22]; the similarity measure between IVNSs N1′ and N2′ is
degrees. Furthermore, Broumi et al. [22] proposed the cor-
defined as follows:
responding similarity measure of IVNSs based on the simi-
larity measure S1SVNS proposed by Majumdar et al. [15].
S1IVNS N1′, N2′􏼁
􏽐ni�1 􏼚min 􏼒TLN′ xi 􏼁, TLN′ xi 􏼁􏼓 + min 􏼒TU

N′
xi 􏼁, TU
N′
xi 􏼁􏼓 + min 􏼒ILN′ xi 􏼁, ILN′ xi 􏼁􏼓 + min 􏼒IU
N′
xi 􏼁, IU
N′
xi 􏼁􏼓 + min 􏼒FLN′ xi 􏼁, FLN′ xi 􏼁􏼓 + min 􏼒FU
N′
xi 􏼁, FU
N′
xi 􏼁􏼓􏼛
1 2 1 2 1 2 1 2 1 2 1 2
� .
􏽐ni�1 􏼚max 􏼒TLN′ xi 􏼁, TLN′ xi 􏼁􏼓 + max 􏼒TU
N′
xi 􏼁, TU
N′
xi 􏼁􏼓 + max 􏼒ILN′ xi 􏼁, ILN′ xi 􏼁􏼓 + max 􏼒IU
N′
xi 􏼁, IU
N′
xi 􏼁􏼓 + max 􏼒FLN′ xi 􏼁, FLN′ xi 􏼁􏼓 + max 􏼒FU
N′
xi 􏼁, FU
N′
xi 􏼁􏼓􏼛
1 2 1 2 1 2 1 2 1 2 1 2
(9)
Remark 3. If TLN′ (xi ) � TU N′j

(xi ), ILN′ (xi ) � IU (x ), FLN′ (xi ) �
N′j i
Theorem 3. The similarity measure S∗1IVNS (N1′, N2′) satisfies
j j j
FU (xi )(j � 1, 2), then the similarity measure S1IVNS (N1′, N2′) the following properties:
N′j
is reduced to the similarity measure S1SVNS (N1 , N2 ). (1) 0 ≤ S∗1IVNS (N1′, N2′) ≤ 1
Similarly to Section 3.1, we propose a corresponding (2) S∗1IVNS (N1′, N2′) � 1 if and only if N1′ � N2′
similarity measure between IVNSs, which is based on the
(3) S∗1IVNS (N1′, N2′) � S∗1IVNS (N2′, N1′)
similarity measure S1IVNS (N1′, N2′) and the Euclidean dis-
tance DIVNS (N1′, N2′) defined in Definition 4.
Proof. The proof is similar to Theorem 1; hence, we omit it
here.
Definition 10. Let X � 􏼈x1 , x2 , . . . , xn 􏼉 be a universal set,
Next, we will use the same method to define the simi-
for any two IVNSs N1′ �{xi , [TLN′ (xi ), TU (xi )], [ILN′ (xi ),
1 N′1 1 larity measure S∗2IVNS (N1′, N2′) between IVNS, which is based
IN′ (xi )], [FN′ (xi ), FN′ (xi )]|xi ∈ X} and N2′ � {xi , [TLN′ }(xi ),
U L U
on the cosine similarity measure S2IVNS (N1′, N2′) proposed
1 1 1 2
TUN′2
(xi )], [ILN′ (xi ), IU
N′2
(xi )], [FLN′ (xi ), FU
N′2
(xi )]|xi ∈ X}; a by Ye [21] (Definition 11) and the Euclidean distance
2 2
new similarity measure S∗1IVNS (N1′, N2′) is defined as DIVNS (N1′, N2′) defined in formula (4). □
follows:
1 any two IVNSs N1′ � {〈xi , [TLN′ (xi ), TU (xi )], [ILN′ (xi ),
S∗1IVNS N1′, N2′􏼁 � S N ′, N ′􏼁 + 1 − DIVNS N1′, N2′􏼁􏼁. N′1
2 1IVNS 1 2 1
IN′ (xi )], [FN′ (xi ), FN′ (xi )]〉|xi ∈ X} and N2′ � {〈xi ,[TLN′
U L U
1
(10) 1
(xi ),TU
1 1
(xi )],[ILN′ (xi ),IU (xi )],[FLN′ (xi ),FU (xi )]〉|xi ∈ X};
2
N′2 2 N′2 2 N′2

The proposed similarity measure also satisfies Theo- the cosine similarity measure S∗2IVNS (N1′,N2′) is defined as
rem 3. follows [21]:
1 n
S2IVNS N1′, N2′􏼁 � 􏽘
n i�1
TLN′ xi 􏼁TLN′ xi 􏼁 + TU
N′1
xi 􏼁TU
N′2
xi 􏼁 + ILN′ xi 􏼁ILN′ xi 􏼁 + IU
N′1
xi 􏼁IU
N′2
xi 􏼁 + FLN′ xi 􏼁FLN′ xi 􏼁 + FU
N′1
xi 􏼁FU
N′2
xi 􏼁
· 􏽲��
2 2
1 2
2 2 2
1 2
2
􏽲��
2
1
2
2
2 2 2 2
.
L U L U
􏼒TN′ xi 􏼁􏼓 + 􏼒TN′ xi 􏼁􏼓 + 􏼒IN′ xi 􏼁􏼓 + 􏼒IN′ xi 􏼁􏼓 + 􏼒FN′ xi 􏼁􏼓 + 􏼒FN′ xi 􏼁􏼓 􏼒TN′ xi 􏼁􏼓 + 􏼒TN′ xi 􏼁􏼓 + 􏼒IN′ xi 􏼁􏼓 + 􏼒IN′ xi 􏼁􏼓 + 􏼒FLN′ xi 􏼁􏼓 + 􏼒FU
L U L U L U
N′
x i 􏼁􏼓
1 1 1 1 1 1 2 2 2 2 2 2
(11)
Example 2. For two IVNSs N1′ � x, [0.3, 0.4], [0.2, 0.3], and the Euclidean distance measure DIVNS (N1′, N2′) as
[0.4, 0.5] and N2′ � x, [0.6, 0.8], [0.4, 0.6], [0.8, 1], according follows:
to formula (11), we have S2IVNS (N1′, N2′) � 1, but N1′ ≠ N2′.
In this case, the necessary condition of (2) in Lemma 1 is
not satisfied. Therefore, the cosine similarity measure
S2IVNS (N1′, N2′) proposed by Ye [22] is not a genuine sim- any two IVNSs N1′ � {〈xi , [TLN′ (xi ), TU N′1
(xi )], [ILN′ (xi ),
1 1
ilarity measure. Motivated by this, we will propose a new IN′ (xi )], [FN′ (xi ), FN′ (xi )]〉|xi ∈ X} and N2′ � {〈xi , [TLN′
U L U
1 1 1 2
similarity measure S∗2IVNS (N1′, N2′) based on S2IVNS (N1′, N2′) (xi ),TUN′
(xi )], [ILN′ (xi ),IU
N′
(xi )], [FLN′ (xi ), FU
N′
(xi )]〉|xi ∈ X};
2 2 2 2 2
a new similarity measure S∗2IVNS (N1′, N2′) can be defined as Table 1: The relation between the diagnosis and the symptom for
follows: SVNS decision information.
1 S1 S2 S3 S4 S5
S∗2IVNS N1′, N2′􏼁 � S2IVNS N1′, N2′􏼁 + 1 − DIVNS N1′, N2′􏼁􏼁.
2 <0.4, 0.6, <0.3, 0.2, <0.1, 0.3, <0.4, 0.3, <0.1, 0.2,
Q1
(12) 0.0> 0.5> 0.7> 0.3> 0.7>
<0.7, 0.3, <0.2, 0.2, <0.0, 0.1, <0.7, 0.3, <0.1, 0.1,
Q2
0.0> 0.6> 0.9> 0.0> 0.8>
Remark 4. In Example 2, when N1′ ≠ N2′, the similarity <0.3, 0.4, <0.6, 0.3, <0.2, 0.1, <0.2, 0.2, <0.1, 0.0,
Q3
measure S2IVNS (N1′, N2′) � 1, this is inconsistent with the real 0.3> 0.1> 0.7> 0.6> 0.9>
decision problems. But, using formula (12) to calculate it again, <0.1, 0.2, <0.2, 0.4, <0.8, 0.2, <0.2, 0.1, <0.2, 0.1,
Q4
0.7> 0.4> 0.0> 0.7> 0.7>
we have S∗2IVNS (N1′, N2′) � 0.8185. Obviously, the proposed
<0.1, 0.1, <0.0, 0.2, <0.2, 0.0, <0.2, 0.0, <0.8, 0.1,
similarity measure S∗2IVNS (N1′, N2′) can rectify the existing Q5
0.8> 0.8> 0.8> 0.8> 0.1>
cosine similarity measure S2IVNS (N1′, N2′) defined by Ye [22].
We can calculate the similarity measures S∗1SVNS (P1 , Qi ) and
S∗2SVNS (P1 , Qi )(i � 1, 2, . . . , 5), and then the diagnoses of the patient P1 can
Theorem 4. The similarity measure S∗2IVNS (N1′, N2′) satisfies be classified by Rj � arg max􏽮S∗jSVNS (P1 , Qi )􏽯(j � 1, 2).
1≤i≤5
the following properties:
(1) 0 ≤ S∗2IVNS (N1′, N2′) ≤ 1
(2) S∗2IVNS (N1′, N2′) � 1 if and only if N1′ � N2′ Table 2: The similarity measures between P1 and Qi .
(3) S∗2IVNS (N1′, N2′) � S∗2IVNS (N2′, N1′) Q1 Q2 Q3 Q4 Q5
S∗1SVNS (P1 , Qi ) 0.6663 0.7188 0.5387 0.4594 0.4336
S∗2SVNS (P1 , Qi ) 0.8223 0.8378 0.6377 0.5500 0.4881
Proof. The proof is similar to Theorem 2, we also omit it
here.
In the next section, we will apply the proposed new are all malaria (Q2 ). The proposed two similarity measures
similarity measures to medical diagnosis decision problem; S∗1SVNS and S∗2SVNS produce the same results as Ye [19], which
numerical examples are also given to illustrate the appli- means the proposed similarity measures are feasible and
cation and effectiveness of the proposed new similarity effective.
measures. □
4. Applications of the Proposed 4.2. The Proposed Similarity Measures between IVNSs for
Similarity Measures Medical Diagnosis. We know if the doctor examines the
patient two or three times a day, then the interval values of
4.1. The Proposed Similarity Measures between SVNSs for multiple inspections for the patient are obtained. In this
Medical Diagnosis. We first give a numerical example about section, we will apply the proposed similarity measures
a medical diagnosis (adapted from Ye [19]) to illustrate the S∗1IVNS and S∗2IVNS to medical diagnosis, the example is also
feasibility of the proposed new similarity measures S∗1SVNS adapted from Ye [19].
and S∗2SVNS between SVNSs.
Example 4. Let us reconsider Example 3, assume a patient
Example 3. Consider a medical diagnosis decision problem; P2 has all the symptoms, which can be expressed by the
suppose a set of diagnoses Q � 􏼈Q1 (viral fever), Q2 following IVNS information.
(malaria), Q3 (typhoid), Q4 (gastritis), Q5 (stenocardia)} and
P2 (Patient) � 􏼈〈S1 , [0.3, 0.5], [0.2, 0.3], [0.4, 0.5]〉,
a set of symptoms S � 􏼈S1 (fever), S2 (headache), S3 (stomach
pain), S4 (cough), S5 (chestpain)}. Assume a patient P1 has all 〈S2 , [0.7, 0.9], [0.1, 0.2], [0.1, 0.2]〉,
the symptoms in the process of diagnosis, the SVNS evaluate 〈S3 , [0.4, 0.6], [0.2, 0.3], [0.3, 0.4]〉, (14)
information about P1 is
〈S4 , [0.3, 0.6], [0.1, 0.3], [0.4, 0.7]〉,
P1 (Patient) � 􏼈〈S1 , 0.8, 0.2, 0.1〉, 〈S2 , 0.6, 0.3, 0.1〉,
〈S5 , [0.5, 0.8], [0.1, 0.4], [0.1, 0.3]〉􏼉.
〈S3 , 0.2, 0.1, 0.8〉, 〈S4 , 0.6, 0.5, 0.1〉, (13)
〈S5 , 0.1, 0.4, 0.6〉􏼉. The same way as Example 3 in Ye [19], the diagnosis
information of SVNSs Qi with respect to symptoms Si (i �
The diagnosis information Qi (i � 1, 2, . . . , 5) with re- 1, 2, · · · , 5) are transformed into IVNSs, which are shown in
spect to symptoms Si (i � 1, 2, . . . , 5) also can be represented Table 3.
by the SVNSs, which is shown in Table 1. By applying formulae (10) and (12), we obtain the
By applying formulae (6) and (8), we can obtain the similarity measure values S∗1IVNS (P2 , Qi ) and S∗2IVNS (P2 , Qi ),
similarity measure values S∗1SVNS (P1 , Qi ) and S∗2SVNS (P1 , Qi ); the results are shown in Table 4.
the results are shown in Table 2. From the two similarity measure values in Table 4, we
From the above two similarity measures S∗1SVNS and can see that the patient P2 suffers from typhoid (Q3 ); the
∗
S2SVNS , we can conclude that the diagnoses of the patient P1 diagnosis results are the same as shown by Ye [19].
Table 3: The relation between the diagnosis and the symptom for IVNS decision information.
S1 S2 S3 S4 S5
<[0.4, 0.4], [0.6, 0.6], <[0.3, 0.3], [0.2, 0.2], <[0.1, 0.1], [0.3, 0.3], <[0.4, 0.4], [0.3, 0.3], <[0.1, 0.1], [0.2, 0.2],
Q1
[0.0, 0.0]> [0.5, 0.5]> [0.7, 0.7]> [0.3, 0.3]> [0.7, 0.7]>
<[0.7, 0.7], [0.3, 0.3], <[0.2, 0.2], [0.2, 0.2], <[0.0, 0.0], [0.1, 0.1], <[0.7, 0.7], [0.3, 0.3], <[0.1, 0.1], [0.1, 0.1],
Q2
[0.0, 0.0]> [0.6, 0.6]> [0.9, 0.9]> [0.0, 0.0]> [0.8, 0.8]>
<[0.3, 0.3], [0.4, 0.4], <[0.6, 0.6], [0.3, 0.3], <[0.2, 0.2], [0.1, 0.1], <[0.2, 0.2], [0.2, 0.2], <[0.1, 0.1], [0.0, 0.0],
Q3
[0.3, 0.3]> [0.1, 0.1]> [0.7, 0.7]> [0.6, 0.6]> [0.9, 0.9]>
<[0.1, 0.1], [0.2, 0.2], <[0.2, 0.2], [0.4, 0.4], <[0.8, 0.8], [0.2, 0.2], <[0.2, 0.2], [0.1, 0.1], <[0.2, 0.2], [0.1, 0.1],
Q4
[0.7, 0.7]> [0.4, 0.4]> [0.0, 0.0]> [0.7, 0.7]> [0.7, 0.7]>
<[0.1, 0.1], [0.1, 0.1], <[0.0, 0.0], [0.2, 0.2], <[0.2, 0.2], [0.0, 0.0], <[0.2, 0.2], [0.0, 0.0], <[0.8, 0.8], [0.1, 0.1],
Q5
[0.8, 0.8]> [0.8, 0.8]> [0.8, 0.8]> [0.8, 0.8]> [0.1, 0.1]>
We can calculate the similarity measures S∗1IVNS (P2 , Qi ) and S∗2IVNS (P2 , Qi )(i � 1, 2, · · · , 5), and the diagnosis of the patient P2 can be classified by
Rj � arg max􏽮S∗jIVNS (P2 , Qi )􏽯(j � 1, 2).
1≤i≤5
Table 4: The similarity measures between P2 and Qi . in X � 􏼈x1 , x2 , . . . , xn 􏼉, the existing similarity measures
Q1 Q2 Q3 Q4 Q5 between N1 and N2 are defined as follows:
S∗1IVNS (P2 , Qi ) 0.5783 0.4610 0.6273 0.5772 0.5401 (1) Broumi et al. [23] proposed the similarity measure
S∗2IVNS (P2 , Qi ) 0.6804 0.5729 0.7503 0.7061 0.6734 SMSVNS :
SMSVNS N1 , N2 􏼁 � 1 − DSVNS N1 , N2 􏼁. (15)
4.3. Comparative Analyses of Existing Similarity Measures.
To illustrate the effectiveness of the proposed similarity (2) Şahin and Ahmet [24] proposed the similarity
measures for medical diagnosis, we will apply the existing measure SDSVNS :
similarity measures of SVNSs and IVNSs for comparative 1
analyses. SDSVNS N1 , N2 􏼁 � . (16)
1 + DSVNS N1 , N2 􏼁
At first, we introduce the existing similarity measures
between SVNSs as follows: (3) Ye [19] proposed the improved cosine similarity
Let N1 � {〈xi , TN1 (xi ), IN1 (xi ), FN1 (xi )〉 ∣ xi ∈ X} and measures SC1SVNS and SC2SVNS :
N2 � {xi , TN2 (xi ), IN2 (xi ), FN2 (xi ) ∣ xi ∈ X} be two SVNSs
􏼌􏼌 􏼌􏼌 􏼌 􏼌 􏼌
π · max􏼒 􏼌􏼌T x 􏼁 − T x 􏼁􏼌􏼌􏼌, 􏼌􏼌􏼌I x 􏼁 − I x 􏼁􏼌􏼌􏼌, 􏼌􏼌􏼌F x 􏼁 − F x 􏼁􏼌􏼌􏼌􏼓
1 n
⎢
⎡
⎢ 􏼌 N i N i 􏼌 􏼌 N i N i 􏼌 􏼌 N i N i 􏼌 ⎥⎥⎤
SC1SVNS N1 , N2 􏼁 � 􏽘 cos⎢
⎢
⎢
1 2 1 2 1 2
⎥⎥⎥, (17)
n i�1 ⎣ 2 ⎦
􏼌􏼌 􏼌 􏼌 􏼌 􏼌 􏼌
π 􏼒 􏼌􏼌T x 􏼁 − T x 􏼁􏼌􏼌􏼌 + 􏼌􏼌􏼌I x 􏼁 − I x 􏼁􏼌􏼌􏼌 + 􏼌􏼌􏼌F x 􏼁 − F x 􏼁􏼌􏼌􏼌􏼓
1 n
⎢
⎡
⎢ 􏼌 N i N i 􏼌 􏼌 N i N i 􏼌 􏼌 N i N i 􏼌 ⎥⎥⎤
SC2SVNS N1 , N2 􏼁 � 􏽘 cos⎢
⎢
⎢
1 2 1 2 1 2
⎥⎥⎥. (18)
n i�1 ⎣ 6 ⎦
(4) Yang et al. [25] proposed the similarity measure Let N1′ � {〈xi , [TLN′ (xi ), TU N′1
(xi )], [ILN′ (xi ), IU
N′1
(xi )],
1 1
SYSVNS (N1 , N2 ): [FN′ (xi ), FN′ (xi )]〉|xi ∈ X} and N2′ � {xi , [TLN′ (xi ), TU
L U
1 1 2 N′2
SCSVNS N1 , N2 􏼁 (xi )], [ILN′ (xi ), IU
N ′ (xi )], [FLN′ (xi ), FU
N ′ (xi )]|xi ∈ X} be two
SYSVNS N1 , N2 􏼁 � . 2 2 2 2
SCSVNS N1 , N2 􏼁 + DSVNS N1 , N2 􏼁 IVNSs in X � 􏼈x1 , x2 , . . . , xn 􏼉, the existing similarity mea-
(19) sures between N1′ and N2′ are defined as follows:
(1) Broumi et al. [23] proposed the similarity measure
Example 3′. We apply formulae (5), (7), and (15)–(19) to SMIVNS :
calculate Example 5 again; the similarity measure values SMIVNS N1′, N2′􏼁 � 1 − DIVNS N1′, N2′􏼁. (20)
between P1 and Qi (i � 1, 2, . . . , 5) are shown in Table 5.
As we can see from Table 5, the patient P1 is still assigned (2) Şahin and Ahmet [24] proposed the similarity
to malaria (Q2 ), and the results are same as the proposed measure SDIVNS :
similarity measures in this paper, which means the proposed
similarity measures are feasible and effective. 1
Next, we introduce the existing similarity measures SDIVNS N1′, N2′􏼁 � . (21)
1 + DIVNS N1′, N2′􏼁
between IVNSs as follows:
Table 5: Comparisons of different similarity measure between Table 6: Comparisons of different similarity measures between
SVNSs. IVNSs.
Q1 Q2 Q3 Q4 Q5 Q1 Q2 Q3 Q4 Q5
SMSVNS [23] 0.7941 0.8094 0.4568 0.5851 0.5517 SMIVNS [23] 0.6833 0.5844 0.7265 0.6923 0.6596
SDSVNS [24] 0.8293 0.8399 0.6480 0.7067 0.6905 SDIVNS [24] 0.7595 0.7064 0.7852 0.7647 0.7460
SC1SVNS [22] 0.8942 0.8976 0.8422 0.6102 0.5607 SC1IVNS [22] 0.7283 0.6079 0.7915 0.7380 0.7157
SC2SVNS [22] 0.9443 0.9571 0.9264 0.8214 0.7650 SC2IVNS [22] 0.8941 0.8459 0.9086 0.9056 0.8797
SYSVNS [25] 0.8128 0.8248 0.6079 0.5953 0.5557 SYIVNS [25] 0.6969 0.5939 0.8429 0.7057 0.6777
S1SVNS [15] 0.5385 0.6282 0.6206 0.3336 0.3154 S1IVNS [22] 0.4733 0.3376 0.5209 0.4620 0.4205
S2SVNS [18] 0.8505 0.8661 0.8185 0.5148 0.4244 S2IVNS [21] 0.6775 0.5613 0.7741 0.7198 0.6872
(3) Broumi and Smarandache [22] proposed the

improved cosine similarity measures SC1SVNS and
SC2SVNS :
1 n π 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌
SC1IVNS N1′, N2′􏼁 � 􏽘 cos 􏼔max􏼒􏼌􏼌􏼌TLN′1 xi 􏼁 − TLN′2 xi 􏼁􏼌􏼌􏼌, 􏼌􏼌􏼌ILN′1 xi 􏼁 − ILN′2 xi 􏼁􏼌􏼌􏼌, 􏼌􏼌􏼌FLN′1 xi 􏼁 − FLN′2 xi 􏼁􏼌􏼌􏼌􏼓
n i�1 4
(22)
􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌
+ max􏼒􏼌􏼌􏼌TU U 􏼌􏼌 􏼌􏼌 U U 􏼌􏼌 􏼌􏼌 U U 􏼌􏼌
N′1 xi 􏼁 − TN′2 xi 􏼁􏼌, 􏼌IN′1 xi 􏼁 − IN′2 xi 􏼁􏼌, 􏼌FN′1 xi 􏼁 − FN′2 xi 􏼁􏼌􏼓􏼕,
1 n π 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌
SC2IVNS N1′, N2′􏼁 � 􏽘 cos 􏼒􏼌􏼌􏼌TLN′1 xi 􏼁 − TLN′2 xi 􏼁􏼌􏼌􏼌 + 􏼌􏼌􏼌ILN′1 xi 􏼁 − ILN′2 xi 􏼁􏼌􏼌􏼌 + 􏼌􏼌􏼌FLN′1 xi 􏼁 − FLN′2 xi 􏼁􏼌􏼌􏼌
n i�1 12
(23)
􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌 􏼌􏼌
+ 􏼌􏼌􏼌TU U 􏼌􏼌 􏼌􏼌 U U 􏼌􏼌 􏼌􏼌 U U
N′1 xi 􏼁 − TN′2 xi 􏼁􏼌 + 􏼌IN′1 xi 􏼁 − IN′2 xi 􏼁􏼌 + 􏼌FN′1 xi 􏼁 − FN′2 xi 􏼁􏼌􏼓.
􏼌􏼌
(4) Yang et al. [25] proposed the similarity measure measure of SVNSs and IVNSs, respectively, which are based
SYSVNS (N1′, N2′): on the above existing similarity measures and the Euclidean
SCIVNS N1′, N2′􏼁 distance measure. And, the similarity measures are proposed
SYIVNS N1′, N2′􏼁 � . not only from the points of view of algebra and geometry but
SCIVNS N1′, N2′􏼁 + DIVNS N1′, N2′􏼁 also satisfy the axiom of the similarity measure. Furthermore,
(24) we apply the proposed similarity measures to medical di-
agnosis decision problems, and the numerical example is used
Example 4′. Applying formulae (9), (11), and (20)–(24) to to illustrate the feasibility and effectiveness of the proposed
calculate Example 6 again, the similarity measure values similarity measure, which are then compared to other existing
between P2 and Qi (i � 1, 2, . . . , 5) are shown in Table 6. similarity measures. In future research, we will focus on
The results of Table 6 show that the patient P2 should be studying the similarity measure between linguistic neu-
assigned to typhoid (Q3 ), they are same as the proposed trosophic set and the application of the proposed similarity
similarity measures S∗1IVNS and S∗2IVNS in the paper, which measures of neutrosophic sets, such as pattern recognition,
means the proposed methods are feasible and effective. supplier selection, and so on.
The proposed similarity measures in the paper have
some advantages in solving multiple criteria decision- Data Availability
making problems. They are constructed based on the
existing similarity measures and Euclidean distance, which The data used to support the findings of this study are in-
not only satisfy the axiom of the similarity measure but also cluded within the article.
consider the similarity measure from the points of view of
algebra and geometry. Furthermore, they can be applied Conflicts of Interest
more widely in the field of decision-making problems.
The authors declare no conflicts of interest regarding the
publication for the paper.
5. Conclusions
Acknowledgments
The similarity measure is widely used in multiple criteria
decision-making problems. This paper proposed a new This research is fully supported by a grant by the National
method to construct the similarity measures combining the Natural Science Foundation of Hunan (2017JJ2096 and
existing cosine similarity measure and the Euclidean distance 2018JJ3137), the National Natural Science Foundation of
China (11501191), the National Social Science Foundation of [18] J. Ye, “Vector similarity measures of simplified neutrosophic
China (15BTJ028), the Outstanding Youth Project of Hunan sets and their application in multicriteria decision making,”
Education Department (1713092), and Major Projects of the International Journal of Fuzzy Systems, vol. 16, no. 2,
National Social Science Foundation of China (17ZDA046). pp. 204–211, 2014.
[19] J. Ye, “Improved cosine similarity measures of simplified
neutrosophic sets for medical diagnoses,” Artificial In-
References telligence In Medicine, vol. 63, no. 3, pp. 171–179, 2015.
[20] D. Liu, X. Chen, and D. Peng, “Interval-valued intuitionistic
[1] L. A. Zadeh, “Fuzzy sets,” Information and Control, vol. 8, fuzzy ordered weighted cosine similarity measure and its
no. 3, pp. 338–353, 1965. application in investment decision-making,” Complexity,
[2] S. K. De, R. Biswas, and A. R. Roy, “Multicriteria decision vol. 2017, article 1891923, 11 pages, 2017.
making using intuitionistic fuzzy set theory,” Journal of Fuzzy [21] D. Liu, X. Chen, and D. Peng, “The intuitionistic fuzzy lin-
Mathematics, vol. 6, no. 4, pp. 591–629, 1998. guistic cosine similarity measure and its application in pattern
[3] N. H. Phuong and V. V. Thang, “Case based reasoning for recognition,” Complexity, vol. 2018, no. 1, pp. 1–11, 2018.
medical diagnosis using fuzzy set theory,” International [22] S. Broumi and F. Smarandache, “New distance and similarity
Journal of Biomedical Soft Computing and Human Sciences, measures of interval neutrosophic sets,” in Proceedings of
vol. 5, pp. 1–7, 2000. International Conference on Information Fusion, pp. 1–7,
[4] O. J. Tobias and R. Seara, “Image segmentation by histogram IEEE, Salamanca, Spain, July 2014.
thresholding using fuzzy sets,” IEEE Transactions on Image [23] S. Broumi and F. Smarandache, “Several similarity measures
Processing, vol. 11, no. 12, pp. 1457–1465, 2002. of neutrosophic sets,” Neutrosophic Sets and Systems, vol. 1,
[5] L. A. Zadeh, “The concept of a linguistic variable and its no. 1, pp. 54–62, 2013.
application to approximate reasoning,” Information Science, [24] R. Şahin and K. Ahmet, On Similarity and Entropy of Neu-
vol. 8, no. 3, pp. 199–249, 1974. trosophic Soft Sets, vol. 29Amsterdam, Netherlands, 2014.
[6] K. T. Atanassov and P. Rangasamy, “Intuitionistic fuzzy sets,” [25] Y. Yang, R. Zhang, and J. Guo, “A multi-attribute decision-
Fuzzy Sets and Systems, vol. 20, no. 1, pp. 87–96, 1986. making approach based on hesitant neutrosophic sets,” Fuzzy
[7] H. Wang, F. Smarandache, Y. Zhang et al., “Single valued Systems and Mathematics, vol. 31, no. 2, pp. 114–122, 2017.
neutrosophic sets,” Review of the Air Force Academy, vol. 3,
no. 1, pp. 33–39, 2013.
[8] H. Wang, “Interval neutrosophic sets and logic: theory and
applications in computing,” Computer Science, vol. 65, no. 4,
pp. 1–99, 2005.
[9] J. Ye, “Multiple attribute group decision-making method
with completely unknown weights based on similarity
measures under single valued neutrosophic environment,”
Journal of Intelligent and Fuzzy Systems, vol. 27, pp. 2927–
2935, 2014.
[10] I. Arockiarani and I. R. Sumathi, “Some results on interval
valued Fuzzy neutrosophic soft set,” International Journal of
Innovative Research and Studies, vol. 3, no. 5, pp. 386–405,
2014.
[11] A. Mukherjee and S. Sarkar, “Several similarity measures of
neutrosophic soft sets and its application in real life prob-
lems,” Annals of Pure and Applied Mathematics, vol. 7, no. 1,
pp. 1–6, 2014.
[12] J. Ye, “Similarity measures between interval neutrosophic sets
and their applications in multicriteria decision-making,”
Journal of Intelligent and Fuzzy Systems, vol. 26, no. 1,
pp. 165–172, 2014.
[13] Beg, Ismat, Ashraf et al., “Similarity measures for fuzzy sets,”
Applied and Computational Mathematics, vol. 8, no. 2,
pp. 192–202, 2009.
[14] Y. Song, X. Wang, and H. Zhang, “A distance measure be-
tween intuitionistic fuzzy belief functions,” Knowledge-Based
Systems, vol. 86, pp. 288–298, 2015.
[15] P. Majumdar and S. K. Samanta, “On similarity and entropy
of neutrosophic sets,” Journal of Intelligent and Fuzzy Systems,
vol. 26, no. 3, pp. 1245–1252, 2013.
[16] J. Ye, “Multicriteria decision-making method based on cosine
similarity measures between interval- valued fuzzy sets with
risk preference,” Economic Computation and Economic Cy-
bernetics Studies and Research, vol. 50, no. 4, pp. 205–215,
2016.
[17] J. Ye, “Cosine similarity measures for intuitionistic fuzzy sets
and their applications,” Mathematical and Computer Mod-
elling, vol. 53, no. 1, pp. 91–97, 2011.
Hindawi
https://doi.org/10.1155/2018/4253928
Research Article
Computational Prediction of the Combined Effect of CRT and
LVAD on Cardiac Electromechanical Delay in LBBB and RBBB
Aulia K. Heikhmakhtiar and Ki M. Lim

Department of IT Convergence Engineering, Kumoh National Institute of Technology, Gumi 39177, Republic of Korea
Correspondence should be addressed to Ki M. Lim; kmlim@kumoh.ac.kr
Received 17 May 2018; Revised 3 September 2018; Accepted 25 October 2018; Published 14 November 2018
Copyright © 2018 Aulia K. Heikhmakhtiar and Ki M. Lim. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
Two case reports showed that the combination of CRTand LVAD benefits the end-stage heart failure patients with prolonged QRS
interval significantly. In one of the reports, the patient had the LVAD removed due to the recovery of the heart function. However,
the quantification of the combined devices has yet to be conducted. This study aimed at computationally predicting the effects of
CRT-only or combined with LVAD on electromechanical behaviour in the failing ventricle with left bundle branch blocked
(LBBB) and right bundle branch blocked (RBBB) conditions. The subjects are normal sinus rhythm, LBBB, RBBB, LBBB with
CRT-only, RBBB with CRT-only, LBBB with CRT + LVAD, and RBBB with CRT + LVAD. The results showed that the CRT-only
shortened the total electrical activation time (EAT) in the LBBB and RBBB conditions by 20.2% and 17.1%, respectively. The CRT-
only reduced the total mechanical activation time (MAT) and electromechanical delay (EMD) of the ventricle under LBBB by
21.3% and 10.1%, respectively. Furthermore, the CRT-only reduced the contractile adenosine triphosphate (ATP) consumption by
5%, increased left ventricular (LV) pressure by 6%, and enhanced cardiac output (CO) by 0.2 L/min under LBBB condition.
However, CRT-only barely affects the ventricle under RBBB condition. Under the LBBB condition, CRT + LVAD increased LV
pressure and CO by 10.5% and by 0.9 L/min, respectively. CRT + LVAD reduced ATP consumption by 15%, shortened the MAT
by 23.4%, and shortened the EMD by 15.2%. In conclusion, we computationally predicted and quantified that the CRT + LVAD
implementation is superior to CRT-only implementation particularly in HF with LBBB condition.
1. Introduction 30% of patients with HF failed to benefit from the CRT [9].
Hu et al. [13] described that three major responses could be
Heart failure (HF) plays a major role in the number of death obtained from the CRT treatment: resynchronization of the
worldwide [1]. Thus, the study of heart diseases including intraventricular contraction, efficient ventricular preloading
cardiac arrhythmia, which progressively leads to HF con- by a properly timed atrial contraction, and reduction of
dition, is very important. Two types of cardiac therapy mitral regurgitation. One of the study findings was that
devices are commonly used to treat patients with cardiac resynchronization of the intraventricular contraction itself
disease: cardiac resynchronization therapy (CRT) and left did not necessarily lead to stroke work improvement.
ventricular assist device (LVAD). CRT is considered as However, the synchronization of the atrioventricular firing
a valuable treatment for patients with dyssynchrony HF with time was essential.
QRS interval >120 ms and left ventricular ejection fraction LVAD supports the ventricular pumping via mechanical
(EF) <35% [2, 3]. A number of studies showed significant unloading for weakened heart. LVAD was initially used as
benefits of using CRT. CRT synchronizes systolic function a bridge to transplantation for patients with end-stage
[4–6], restores heart structure [7, 8], and improves symp- disease [14, 15]. However, because the availability of heart
toms and the quality of life of the patients identified by the donors is very limited, LVAD is now used as destination
improvement of exercise endurance [3, 8–12]. However, therapy as it lasts for years [16–18]. LVAD reverses the
damaged ventricle and recovers myocardial functionalities 2. Materials and Methods

by repairing left ventricular (LV) mass [19, 20], heart
chamber size [20, 21], improves mitral filling [22], and We followed an existing well-developed electromechanical
induces cardiomyocyte hypertrophy regression [21, 22]. The failing ventricle model with the fibers and laminar sheet
use of LVAD also increased the quality of life of the patients structures based on diffusion tensor magnetic resonance
[23–25]. Previously, we observed the effect of LVAD on the imaging [34–37]. In this study, our electromechanical model
electromechanical delay (EMD) under mild and severe HF was coupled with circulatory systems and LVAD models
conditions [26]. The results showed that the LVAD not only similarly as Lim et al. with additional Purkinje networks
increased the EF but also shortened the EMD under the mild compartment to simulate LBBB and RBBB conditions [38].
HF and even better under severe HF conditions. The The electromechanical model consisted of electrophysio-
comparison study of the symptomatic relief of CRT and logical and myofilament dynamics model coupled by Ca2+
LVAD by Delgado et al. concluded that the use of both transient. The LVAD function included the circulatory
devices could synergistically improve cardiac functions for systems, which connected to the 3D ventricular mechanics.
severe HF treatment [27]. Figure 1 shows a full schematic of the system we used in this
In 2011, a case report article stated that the CRT study.
supported the restoration of end-stage HF in a 15-year-old The electrical component was a failing ventricular mesh
patient who underwent LVAD implantation [28]. The CRT with 241,725 nodes and 1,298,751 tetrahedron elements.
shortened the septal to posterior wall motion delay from The electrical mesh has the characteristics of realistic heart
146 ms to 104 ms, and overall, it backed up the hemody- compartment, which consists of endocardial, mid-
namic improvements. After such improvement, the pa- myocardial, and epicardial cells following the ten Tusscher
tient successfully had the LVAD removed. In another case et al. human ventricular tissue model [39]. The line mesh
report article, the CRT and LVAD cooperatively restored type representing Purkinje networks was mapped onto
the cardiac functions of a 62-year-old patient who had a 3D ventricle chamber as well at the endocardial region
cardiogenic shock and left bundle branch block (LBBB) (Figure 2(c)).
[29]. The report stated that the combination of CRT and The Purkinje networks induced the electrical activation
LVAD performed a profound treatment on the weakening sequences of sinus rhythm (normal), LBBB, and RBBB
heart under LBBB condition. It recovered the heart hy- (Figure 2(c)). The CRT pacing site of the LBBB was placed at
pertrophy, and the EF was increased. Based on these re- the LV free-wall, while the CRT pacing site of the RBBB was
ports, we conducted a computational modelling which placed at the RV endocardial apex as shown in Figure 2(b).
combined the CRT and LVAD to the failing ventricle with The electrical stimulation was first induced at the root node
LBBB and RBBB in order to understand the mechanism of of Purkinje fiber model, which propagates to the terminals,
the two devices combined to provide such improvement to hence stimulating the ventricular tissue. The electrical
the heart. propagation in the Purkinje can be described by solving
To the best of our knowledge, a computational study a one-dimensional wave propagation equation and triggered
that quantifies the effect of combined CRT and LVAD has the ventricular activation [40].
yet to be conducted. This study aims at predicting, com- The electrical propagation signal represents an ion ex-
putationally with an electromechanical failing ventricles change across the myocyte as described by ten Tusscher et al.
model, the effects of CRT alone and the combination of [39]. The electrophysiological phenomenon for the single
CRT and LVAD treatment in patients with LBBB or right cell can be described as follows:
bundle branch block (RBBB). We used a well-developed
electromechanical ventricular model which had been used dV −Iion + Istim
to observe different heart conditions, including the pro- � , (1)
dt Cm
longed EMD in the dyssynchrony HF condition [30],
mechanoelectrical feedback on scroll wave stability [31], where V represents the voltage difference of intracellular and
and spontaneous arrhythmia in acute regional ischemia extracellular, t represents time, Iion represents the sum of all
[32]. Recently, by using similar electromechanical cou- transmembrane ionic currents, Istim represents the current if
pling model, our group revealed the effect of KCNQ1 an external stimulus is applied, and Cm represents the cell
S140G mutation on arrhythmogenesis and pumping capacitance per unit of surface area. Iion consists of major
performance [33]. We used computational methods to ionic current as follows:
overcome the measurement limitations and risks of the
experimental study. We analyzed seven HF diseases and
Iion � INa + IK1 + Ito + IKr + IKs + ICaL + INaCa + INaK
therapies: (i) normal sinus rhythm, (ii) LBBB, (iii) LBBB
with CRT alone (LBBB + CRT), (iv) LBBB with CRT and + IpCa + IpK + IbCa + IbNa ,
LVAD combination (LBBB + CRT + LVAD), (v) RBBB, (2)
(vi) RBBB with CRT alone (RBBB + CRT), and (vii) RBBB
with combined CRT and LVAD (RBBB + CRT + LVAD) where INa is fast inward Na+ current, IK1 is inward rectifier
conditions. K1 current, Ito is transient outward K+ current, IKr is rapid
Electrical component Mechanical component
RPA RPV CLA
CPA CPV RMI
Ipk
Ito
INa,K PRV VLV
RPU
CRV
Ip,Ca
CLV
INa,Ca VRV PLV
RAO
LVAD
EK Ikl
EK IKs
RTR RSV RSA
EK IKr
Extracellular
ECa ICa,L CRA CSV CSA
Iup Ca2+
SR ECa ICa,b
Total
ECa force
ECa Ileak ICa,L NXB
Irel Intracellular
ENa INa,b Series
ECa knp(TCaTot)7.5 knp(TCaTot)–7.5 elastic
ENa INa Mass
ECa Ixfer PXB
Cm gxbT fappT
gappT Passive Active
force force Viscosity
hfT
XBPostR XBPreR
hbT
Figure 1: Schematics of the electrical and mechanical elements coupled with transient calcium. Electrical element: it represents the currents,
pumps, and exchanger of the ten Tusscher ionic model as explained in equation (1). Mechanical element: a schematic diagram of the finite-
element ventricular mechanical model coupled with the circulatory and LVAD models. PRV, RV pressure; VRV, RV volume; PLV, LV
pressure; VLV, LV volume; RPA, pulmonary artery resistance; CPA, pulmonary artery compliance; RPV, pulmonary vein resistance; CPV,
pulmonary vein compliance; RMI, mitral valve resistance; CLA, left atrium compliance; RAO, aortic valve resistance; RSA, systemic artery
resistance; CSA, systemic artery compliance; RSV, systemic vein resistance; CSV, systemic vein compliance; RTR tricuspid valve resistance;
CRA, right atrium compliance; RPU, pulmonary valve resistance.
delayed rectifier K+ current, IKs is slow delayed rectifier K+ Ileak � Vleak Casr − Cai 􏼁,
current, ICaL is L-type inward Ca2+ current, INaCa is Na+/Ca2+
exchanger current, INaK is the Na+/K+ pump current, IpCa Vmaxup
and IpK are sarcoplasmic plateau Ca2+ and K+ currents, IbCa Iup � ,
1 + K2up /Ca2i
is background Ca2+ current, and INa,b is background Na+
current.
Ca2sr
To represent the electrical propagation through the Irel � 􏼠arel + Crel 􏼡 dg,
conduction in 3D spatial, the cardiac tissue in this case could b2rel + Ca2sr
be described by the combination of the Equation (1) with the
cellular resistivity (ρ) and surface-to-volume ratio (S) in x, y, Cai × Buf c
Caibufc � ,
and z directions, respectively. This phenomenon can be Cai + Kbufc
described by the following partial differential equation:
dCaitotal −ICal + IbCa + IpCa − 2INaCa
� + Ileak − Iup + Irel ,
dV −Iion + Istim 1 zV 2 dt 2Vc F
� +
dt Cm ρx Sx Cm zx2 Casr × Buf sr
(3) Casrbufsr � ,
2 2 Casr + Kbufsr
1 zV 1 zV
+ + .
ρy Sy Cm zy ρz Sz Cm zz2
2
dCastotal VC
� 􏼐−Ileak + Iup − Irel 􏼑,
dt Vsr
The mathematical equation for calcium dynamics for (4)
coupling the electromechanical simulation was also de- where Ileak represents the calcium released from the sar-
scribed by ten Tusscher et al. [39]: coplasmic reticulum (SR) into the cell. Iup represents the
Endo M Epi
LBBB pacing
site
RBBB
pacing site
(a) (b)
Sinus LBBB RBBB
19 120 (ms)
(c)
Figure 2: (a) The heterogeneous mesh that has the characteristics of endocardial, midmyocardial, and epicardial cells. (b) The CRT pacing
site for the LBBB heart was placed on the LV free-wall, while that for the RBBB heart was placed inside the RV at the bottom of the septum.
(c) The electrical propagation by the Purkinje network in the sinus, LBBB, and RBBB conditions. The electrical activation by the Purkinje
networks is indicated in red. In sinus pacing, the Purkinje network delivers the electrical signal to its terminals. In the LBBB condition, the
Purkinje network delivered the signal only to the right network in the RV area. In the RBBB condition, the Purkinje network delivered the
signal only to the left network in the LV area. CRT, cardiac resynchronization therapy; LBBB, left bundle branch block; LV, left ventricular;
RBBB, right bundle branch block; RV, right ventricular.
calcium pumping to restore the calcium again back to the tricular contraction results from the active tension generation
SR. Irel represents the calcium-induced calcium release represented by the myofilament dynamics model described by
current, d is the activation gate of Irel , which was the same Rice et al. [41]. Ventricular deformation is described by the
with ICaL . Caitotal is the total calcium inside the cell, which equations of passive cardiac mechanics, with the myocardium
consists of Caibufc , the calcium buffer inside the cell, and being orthotropically hyperelastic, and nearly incompressible
Cai , the free calcium inside the cell. Accordingly, the material with passive properties defined by an exponential
Casrtotal is the sum of calcium in the SR, which includes strain-energy function. Simultaneous solutions of the myo-
Casrbufsr , the calcium buffer, and Casr , the free calcium filament model equations to those representing passive car-
inside the SR. diac mechanics on the finite-element mesh constitute cardiac
A Ca2+ transient serves as an input to the cell myofilament contraction. Considering the isometric contraction, we as-
model representing the generation of active tension within sumed that the sarcomere length (SL) is 0 at the initial value,
each myocyte in which an ODE set and multiple algebraic dSL/dt � 0. To measure the isotonic contraction, the SL is
equations describe Ca2+ binding to troponin C, cooperatively solved using the following ordinary differential equation
between regulatory proteins, and cross-bridge cycling. Ven- (ODE):
d IntegralForce + SL0 − SL􏼁 × viscosity

(5)
3. Results and Discussion
SL � ,
dt mass
3.1. Electrophysiological Simulation Results. Figure 3 shows
the viscosity and mass are described in Figure 1 at the the membrane potential propagation in the normal sinus
mechanical compartment, and IntegralForce is the sum of the rhythm (sinus), LBBB, LBBB + CRT, LBBB + CRT + LVAD,
normalized force integrated toward time: RBBB, RBBB + CRT, and RBBB + CRT + LVAD conditions.
t The electrical wavelength almost covered the whole ventricle
IntegralForce � 􏽚 Factive (x) + Fpassive (x) tissue at 100 ms in the sinus condition, and the EAT in the
0 (6) sinus condition was 120 ms (Table 1). The EAT in the LBBB
− Fpreload − Fafterload (x)􏼁 dt, condition was 173 ms, which was the longest among other
cases. CRT shortened the EAT in the LBBB condition to
where the Factive (x) is the active force and the Fpassive (x) is 138 ms, which was close to that in the sinus condition. In the
the passive force. Fpreload is the constant force at the resting RBBB heart, the EAT was 164 ms. CRT shortened the EAT in
length of the initial sarcomere length, and Fafterload is the the RBBB heart to 136 ms, which was close to the sinus
force during the isotonic or isometric contraction. Thus, condition as well.
Fafterload is expressed as follows: Figure 4 shows the calcium activation sequences. The
Fafterload (x) � KSE × x − SL0 􏼁. (7) calcium activation sequence followed the electrical activa-
tion sequence for each case. Compared to the membrane
We calculated the ATP consumption by integrating the potential, which deactivated after 450 ms, the Ca2+ de-
myofilament model in one cycle proposed by Rice et al. from activation occurred at 250 ms. CRT fastened the Ca2+ acti-
each node [41]. In the myofilament model, the ATP con- vation throughout the ventricle in the LBBB condition but
sumption rate (E) is the outcome of cross-bridge detachment insignificantly in the RBBB condition due to the CRT pacing
rate (gxbT ) and the single overlap fraction of thick filaments site.
(SOVFThick):
E � gxbT × SOVFThick . (8)
3.2. Cardiac Mechanics Simulation Results. Figure 5 shows
To construct an integrated model of an LVAD- the 3D ATP consumption rate, tension, and strain trans-
implanted cardiovascular system, we added a compart- mural distribution in the sinus, LBBB, LBBB + CRT, LBBB +
ment of LVAD function between LV and systemic arteries in CRT + LVAD, RBBB, RBBB + CRT, and RBBB + CRT +
the circulatory system based on Kerckhoffs et al. [42] as LVAD conditions. We pick one node at the LV free-wall as
described previously by Lim et al. [38]. Briefly, the LVAD representative to compare the values of them. Overall, the
component was modelled as a flow generator with a specific ATP consumption rate and tension in the LBBB condition
mean flow rate of 3 L/min. Constant-flow conditions were had the highest values, 3.3 and 3.5 times larger than those in
used to simulate the continuous LVAD with the inlet at the the normal sinus condition, respectively. However, the CRT
LV apex and the outlet at the ascending aorta. decreased the ATP consumption rate and tension in the
For the simulation protocol, first we simulated the LBBB condition to be 6% and 7% lower than those in the
electrical model with the Purkinje delivering the signal normal sinus condition. Furthermore, CRT and LVAD re-
representing sinus rhythm, LBBB, or RBBB until the steady duced the ATP consumption rate and tension in the LBBB
state was achieved. We set the conduction velocity by condition to 10% and 12% lower than those in the normal
60 cm/s and the basic cycle length by 600 ms. Here, we sinus condition, respectively. The ATP consumption rate
incorporated the dyssynchrony HF conditions (LBBB and and tension in the RBBB condition were 8% and 9% higher
RBBB) using CRT-alone or combined CRT and LVAD. We than those in the control condition. The CRT reduced the
then used Gaussian point for the interpolation of transient ATP consumption rate and tension by 4% and 5%, re-
Ca2+ from the electrical simulation as the input to the spectively. CRT and LVAD reduced the ATP consumption
mechanical simulation. To model the HF condition, we rate and tension by 9% and 11% in the RBBB condition,
multiplied the constant of passive scaling in the strain- respectively. The strain under the LBBB condition was 80
energy function by 5 times to stiffen the myocardium. The times larger than that in the normal condition (notified by
mechanical model was simulated for 20 seconds to reach major red region in the LV free-wall). The CRT reduced the
the steady state. We compared the ATP consumption rates strain in the LBBB 56 times lower than that in the normal
and tension activation during end-systolic volume (ESV) sinus condition. CRT significantly restored the total strain
and the strain during end-diastolic volume (EDV) by in- under the LBBB condition with the pacing site at the LV
tegrating the information of them from each node. EMD free-wall. However, under LBBB + CRT + LVAD, the total
was defined as the time interval between mechanical ac- strain was only eight times lower than that in the sinus
tivation time (MAT) and electrical activation time (EAT). condition. The continuous LVAD altered the overall strain
MAT was identified when the local strain was shortened to in the ventricles. The total strain in the RBBB condition was
10% before its maximum, while EAT was identified when 84 times larger than that in the normal condition. In the
the myocyte started to depolarize, which in our case was RBBB + CRT model, the strain activation was 122 times
−50 mV [30]. larger. In the RBBB + CRT + LVAD model, the strain
Sinus
LBBB
LBBB + CRT
(+LVAD)
RBBB
RBBB + CRT
(+LVAD)
50 ms 100 ms 150 ms 400 ms 450 ms 500 ms

–85 30
Membrane potential (mV)
Figure 3: The membrane potential propagation of the sinus pacing, LBBB, LBBB + CRT, RBBB, and RBBB + CRT models. For the LBBB +
CRT + LVAD and RBBB + CRT + LVAD models, we used the same electrical activation sequence as the LBBB + CRT and RBBB + CRT
models, respectively, because the LVAD was incorporated into the mechanical computation. LBBB, left bundle branch block; CRT, cardiac
resynchronization therapy; RBBB, right bundle branch block; LVAD, left ventricular assist device.
Table 1: Hemodynamic responses under normal, LBBB, LBBB + CRT, LBBB + CRT + LVAD, RBBB, RBBB + CRT, and RBBB + CRT +
LVAD models.
Condition EDV (mL) ESV (mL) CO (L/min) EF (%) Longest EAT (ms) Average MAT (ms) Average EMD (ms)
Normal sinus rhythm 88 55 3.4 38 120 157 78
LBBB 90 60 3 33.4 173 188 79
LBBB + CRT 89 57 3.2 36 138 148 71
LBBB + CRT + LVAD — — 3.9 — 138 144 67
RBBB 87 54 3.4 38 164 162 80
RBBB + CRT 88 55 3.4 38 136 157 81
RBBB + CRT + LVAD — — 4 — 136 155 79
activation was 23 times lower than that in the normal did not quantify the EF in models in which LVAD model was
condition. incorporated because the PV-loop of the LV was altered due
Figure 6(a) shows the LV pressure-volume (PV) loop to the LVAD implementation. However, the CO was pre-
diagram in the sinus, LBBB, LBBB + CRT, LBBB + CRT + sented to compare the effects of CRT-alone and combined
LVAD, RBBB, RBBB + CRT, and RBBB + CRT + LVAD CRT and LVAD. The complete EDV, ESV, CO, EF, MAT,
conditions. The LV PV-loop in the normal sinus rhythm and EMD data in all conditions are provided in Table 1.
condition was the same as those in the RBBB and RBBB + Figure 6(b) shows the LV and systemic artery pressures
CRTconditions. The EDV in the three conditions was 88 mL, coloured in black and red lines, respectively. In the normal
ESV was 54.5 mL, and EF was 38%. The LBBB condition had sinus condition, the peak LV pressure was 148 mmHg. The
the highest EDV and ESV of 90 mL and 60 mL, respectively. LV pressure peaked at 358 ms, and the aortic valve was
The stroke volume (SV) and EF were 30 mL and 33.4%, opened for 96 ms. In the LBBB condition, the LV peak
respectively. CRT increased the EF up to 36% in the LBBB pressure was 132.5 mmHg. The LV peak pressure time was
condition. The EDV, ESV, and SV in the LBBB + CRT prolonged to 407 ms, and the aortic valve was opened for
condition were 89 mL, 57 mL, and 32 mL, respectively. We 100 ms. The LV peak pressure in the LBBB + CRT condition
Sinus
LBBB
LBBB + CRT
(+LVAD)
RBBB
RBBB + CRT
(+LVAD)
50 ms 100 ms 150 ms 250 ms 350 ms 500 ms
0 0.0008
Calcium
Figure 4: The Ca2+ propagation sequence following the membrane potential activation sequence in all cases. LBBB, left bundle branch block;
CRT, cardiac resynchronization therapy; LVAD, left ventricular assist device; RBBB, right bundle branch block.
5 × 10–4
ATP
(s–1)
0
30
(kPa)
Tension
0
0.15
Strain
–0.1
Sinus LBBB LBBB + CRT LBBB + CRT RBBB RBBB + CRT RBBB + CRT
+ LVAD + LVAD
Figure 5: The 3D transmural distribution of ATP, tension, and strain in all cases. The snapshot for the ATP and tension was taken at the end-
systolic volume, while the strain snapshot was taken at the end-diastolic volume time. ATP, adenosine triphosphate; LBBB, left bundle
branch block; CRT, cardiac resynchronization therapy; LVAD, left ventricular assist device; RBBB, right bundle branch block.
was increased to 141 mmHg. The LV peak pressure time was 148 mmHg and 353 ms, respectively, and the aortic valve was
shortened to 364 ms, and the aortic valve was opened for opened for 96 ms, the same as that in the normal condition.
96 ms. In the LBBB + CRT + LVAD condition, the peak LV The reason was that the electrical activation of the LV was
pressure was 148 mmHg, the same as that in the normal not altered in the RBBB condition. In the RBBB + CRT +
condition, the LV peak pressure time was 363 ms, and the LVAD condition, the peak LV pressure and LV peak
aortic valve was opened for 62 ms. The combination of CRT pressure time were increased to 155 mmHg and 350 ms,
and LVAD restored the LV peak pressure and LV peak respectively, and the aortic valve was opened for 63 ms. This
pressure time in the LBBB condition better than CRT alone. finding shows that the combined CRT and LVAD increased
In the RBBB and RBBB + CRT conditions, the LV peak the LV pressure and systemic artery more than the CRT-only
pressures and LV peak pressure time were the same at implementation.
180 180
150
135
Pressure (mmHg)
Pressure (mmHg)
120
90 90
60
45
30
0 0
0 20 40 60 80 100
Sinus
LBBB
LBBB + CRT
LBBB + CRT
RBBB
RBBB + CRT
RBBB + CRT
+ LVAD
+ LVAD
Volume (mL)
Sinus RBBB + CRT
LBBB LBBB + CRB + LVAD
RBBB RBBB + CRT + LVAD
LBBB + CRT
(a) (b)
120 5
100 4
Cardiac output (L/min)
80
3
ATP
60
2
40
1
20
0 0
Sinus LBBB RBBB Sinus LBBB RBBB
+CRT +CRT
+CRT + LVAD +CRT + LVAD
Control Control
(c) (d)
Figure 6: (a) LV pressure-volume loop, (b) LV pressure and aortic pressure, (c) ATP consumption rate, and (d) cardiac output of all cases.
LV, left ventricular; LBBB, left bundle branch block; CRT, cardiac resynchronization therapy; LVAD, left ventricular assist device; RBBB,
right bundle branch block; ATP, adenosine triphosphate.
Figure 6(c) shows the overall ATP consumption rate cases. As shown in Figure 7(a), the activation sequence of
from one cycle of steady state. In the normal sinus condition, MAT and EMD was identical to the EAT. The MAT and EMD
the ATP consumption rate was 93 s−1. The ATP con- values in the normal condition were 157 and 78 ms, re-
sumption rate in LBBB condition was the highest at 100 s−1. spectively. In the LBBB condition, the MAT and EMD values
CRT reduced the ATP consumption rate in the LBBB were the greatest at 188 ms and 79 ms, respectively. CRT
condition by 5%, while CRT and LVAD reduced the ATP shortened the MAT and EMD to 148 ms and 71 ms in the
consumption rate by 15%. In the RBBB and RBBB + CRT LBBB condition. Furthermore, the MAT and EMD in the
conditions, the ATP consumption rates were the same as LBBB condition were further shortened with the combined
that in the normal condition, 93 s−1. However, with CRT and CRT and LVAD to 144 and 67 ms, respectively. This finding
LVAD support, the ATP consumption rate in the RBBB showed that combination of CRT and LVAD performed
condition was decreased by 16% to 84 s−1. better than CRT alone despite having the same EAT. In the
Figure 6(d) shows the CO of the seven subjects. CRT RBBB condition, the MAT and EMD values were 162 and
increased the CO slightly in the LBBB condition but not in 80 ms, respectively. In the RBBB + CRT condition, the mean
the RBBB condition. With the combination of CRT and MATand EMD values were reduced slightly to 157 and 81 ms,
LVAD, the CO of the LBBB and RBBB conditions was respectively. In the RBBB + CRT + LVAD condition, the
significantly increased (Table 1). mean MAT and EMD values were reduced to 155 and 79 ms
Figure 7(a) shows the 3D contour of EAT, MAT, and (close to the control condition), respectively. In the RBBB
EMD, while Figure 7(b) shows the MAT and EMD of all condition, CRT alone only slightly affected the electrical and
100
(ms)
EAT
50
200
(ms)
MAT
80
115
(ms)
EMD
30
Sinus LBBB LBBB + CRT LBBB + CRT RBBB RBBB + CRT RBBB + CRT
+ LVAD + LVAD
(a)
200
Avg. MAT and EMD (ms)
150
100
50
0
Sinus
LBBB
LBBB + CRT
LBBB + CRT
RBBB
RBBB + CRT
RBBB + CRT
+ LVAD
+ LVAD
MAT
EMD
(b)
Figure 7: (a) Three-dimensional contour distribution of the EAT, MAT, and EMD and (b) the mean MATand EMD values in all cases. EAT,
electrical activation time; MAT, mechanical activation time; EMD, electromechanical delay; LBBB, left bundle branch block; CRT, cardiac
resynchronization therapy; LVAD, left ventricular assist device; RBBB, right bundle branch block.
mechanical responses. However, CRT + LVAD increased CO (3) CRT reduced the ATP consumption by 5% as well as
in the RBBB condition significantly as expected. the tension and strain in the LBBB condition. CRT
In general, we performed a simulation and analyzed also slightly increased the LV peak pressure by 6%
seven cardiac diseases and therapy conditions by using an and increased the CO by 0.2 L/min in the LBBB
electromechanical ventricular model incorporated with condition. However, CRT-alone did not affect these
a circulatory systems, CRT, and LVAD models. The models mechanical responses in the RBBB condition.
including normal sinus rhythm, LBBB, LBBB + CRT, LBBB (4) Combination of CRT and LVAD reduced the ATP
+ CRT + LVAD, RBBB, RBBB + CRT, and RBBB + CRT + consumption by 15% in the LBBB condition and by
LVAD. The major findings of this study are as follows: 16% in the RBBB condition. It also increased the LV
(1) CRT shortened the longest EAT by 20.2% in the pressure by 10.5% in the LBBB condition and 5.7% in
LBBB condition and 17.1% in the RBBB condition. the RBBB condition as well as the CO up to 4 L/min
CRT shortened EMD by 10.1% in the LBBB condi- for both conditions, a degree greater than that in the
tion but did not shorten EMD in the RBBB control condition.
condition. The pacing site at the tissue greatly affected synchro-
(2) Combination of CRT and LVAD treatment short- nization. Placement at the LV free-wall showed faster ac-
ened EMD more than CRT alone in the LBBB tivation throughout the chamber since the signal propagated
condition (15.2%) and shortened EMD in the RBBB evenly from the midway between the base and apex
condition (1.3%). throughout the LV tissue. However, the pacing site at the RV
endocardial apex showed a longer activation time for the the MAT and EMD more than CRT-alone, restored the
electrical signal to propagate to the base. hemodynamic, and produced a greater CO than normal in
In the RBBB condition, the electrical activation of the LV the LBBB and RBBB conditions. Using the combined system,
chamber was the same as normal sinus despite the electrical LVAD contributed to the MAT reduction by mechanical
activation alteration in the RV. This is the major factor why unloading, shortened the EMD, reduced ATP consumption,
the mechanical responses of the LV (LV PV-loop, LV and reduced tension, which contributes to the recovery of
pressure, and CO) were also the same as that in normal sinus the heart shape and function. In short, we computationally
rhythm. Even the CRT implementation in the RBBB con- predicted and quantified that the CRT + LVAD imple-
dition did not affect the mechanical responses of the LV mentation is superior to CRT-only implementation par-
chamber. However, we observed significant improvement in ticularly in HF with the LBBB condition.
CO in the RBBB + CRT + LVAD condition. CRT and LVAD
also reduced the ATP consumption and tension and in- Data Availability
creased the LV and aortic pressure in the RBBB condition.
CRT-alone did not fully restore the mechanical re- The methods and results data used to support the findings of
sponses under the LBBB condition to normal despite the this study are included within the article.
resynchronization of the EAT, which shortened the MAT
and EMD to back normal. On the contrary, the use of Conflicts of Interest
combined CRT and LVAD reduced the energy consumption
(indicated by the ATP consumption rate), tension, increased The authors declare that they have no conflicts of interest.
the LV pressure, and produced CO by 18% more than that in
the HF with sinus rhythm condition. In addition, the LVAD Acknowledgments
did not fully assist the blood distribution in the LBBB or
RBBB condition. This research was partially supported by the MSIT (Ministry
The computational method allowed us to predict the of Science, ICT), under the ITRC support program (IITP-
electromechanical phenomenon in the dyssynchrony heart 2018-2014-0-00639) and and Global IT Talent Support
which underwent CRT and LVAD treatment. It is hardly Program (IITP-2017-0-01811) supervised by the IITP, and
possible to observe some parameters including ATP con- NRF under basic engineering research project
sumption, tension, MAT, and EMD from the patient’s heart (2016R1D1A1B0101440) and the EDISON (NRF-2011-
of that condition in experimental procedure. This study 0020576) Programs.
quantified the electrical activation and hemodynamic re-
sponses in several dyssynchrony HF combined with CRT- References
only, and CRT and LVAD, which has never been conducted
[1] D. Lloyd-Jones, R. Adams, M. Carnethon et al., “Heart disease
previously. As for the clinical impact, the results of this study
and stroke statistics—2009 update: a report from the
can be used as reference to generally predict the effect of the American Heart Association Statistics Committee and Stroke
combination of CRT and LVAD devices to the HF patients Statistics Subcommittee,” Circulation, vol. 119, no. 3,
with LBBB. Though, there are some parameters need to be pp. e21–e181, 2009.
consider deeply by the cardiologist expert. [2] C. Linde, C. Leclercq, S. Rex et al., “Long-term benefits of
This computational study has several limitations that biventricular pacing in congestive heart failure: results from
need to be addressed. The study did not follow the standard the MUltisite STimulation in cardiomyopathy (MUSTIC)
biventricular pacing method [43]. Instead, we used only one study,” Journal of the American College of Cardiology, vol. 40,
CRT pacing site: the RV endocardial apex (for the RBBB no. 1, pp. 111–118, 2002.
condition) or the LV lateral wall (for the LBBB condition). [3] J. B. Young, W. T. Abraham, A. L. Smith et al., “Combined
cardiac resynchronization and implantable cardioversion
The LV pacing site was not placed at the optimal position as
defibrillation in advanced chronic heart failure: the MIRACLE
described before [44]. We did not observe or describe the RV ICD Trial,” JAMA, vol. 289, no. 20, pp. 2685–2694, 2003.
mechanical responses in the RBBB condition; instead, we [4] M. Kawaguchi, T. Murabayashi, B. J. Fetics et al., “Quanti-
showed the LV responses only. We used LVAD instead of tation of basal dyssynchrony and acute resynchronization
a right ventricular assist device to support the RBBB con- from left or biventricular pacing by novel echo-contrast
dition. Because of our limitations, we did not validate the variability imaging,” Journal of the American College of
results of our simulation with experimental data. In addi- Cardiology, vol. 39, no. 12, pp. 2052–2058, 2002.
tion, we did not describe long-term effects such as recovery [5] G. S. Nelson, R. D. Berger, B. J. Fetics et al., “Left ventricular or
of cardiac functions by the combined CRT and LVAD, as biventricular pacing improves cardiac function at diminished
previously described [28, 29]. energy cost in patients with dilated cardiomyopathy and left
bundle-branch block,” Circulation, vol. 102, no. 25,
pp. 3053–3059, 2000.
4. Conclusions [6] P. Søgaard, W. Y. Kim, H. K. Jensen et al., “Impact of acute
biventricular pacing on left ventricular performance and
In conclusion, although CRT-alone shortened the MAT and volumes in patients with severe heart failure,” Cardiology,
EMD to more than normal in the LBBB condition, the vol. 95, no. 4, pp. 173–182, 2001.
mechanical responses in the LBBB condition were not re- [7] A. Auricchio, C. Stellbrink, S. Sack et al., “Long-term
stored to normal. The combined CRT and LVAD shortened clinical effect of hemodynamically optimized cardiac
resynchronization therapy in patients with heart failure and of chronic unloading by an implantable ventricular assist
ventricular conduction delay,” Journal of the American device,” Journal of the American College of Cardiology, vol. 27,
College of Cardiology, vol. 39, no. 12, pp. 2026–2033, 2002. no. 4, pp. 894–901, 1996.
[8] M. G. S. J. Sutton, T. Plappert, K. E. Hilpisch et al., “Sustained [23] K. L. Grady, P. Meyer, A. Mattea et al., “Improvement in
reverse left ventricular structural remodeling with cardiac quality of life outcomes 2 weeks after left ventricular assist
resynchronization at one year is a function of etiology device implantation,” Journal of Heart and Lung Trans-
quantitative doppler echocardiographic evidence from the plantation, vol. 20, no. 6, pp. 657–669, 2001.
Multicenter InSync Randomized Clinical Evaluation (MIR- [24] F. D. Pagani, L. W. Miller, S. D. Russell et al., “Extended
ACLE),” Circulation, vol. 113, no. 2, pp. 266–272, 2006. mechanical circulatory support with a continuous-flow rotary
[9] W. T. Abraham, W. G. Fisher, A. L. Smith et al., “Cardiac left ventricular assist device,” Journal of the American College
resynchronization in chronic heart failure,” New England of Cardiology, vol. 54, no. 4, pp. 312–321, 2009.
Journal of Medicine, vol. 346, no. 24, pp. 1845–1853, 2002. [25] M. S. Slaughter, J. G. Rogers, C. A. Milano et al., “Advanced
[10] S. Cazeau, C. Leclercq, T. Lavergne et al., “Effects of multisite heart failure treated with continuous-flow left ventricular
biventricular pacing in patients with heart failure and in- assist device,” New England Journal of Medicine, vol. 361,
traventricular conduction delay,” New England Journal of no. 23, pp. 2241–2251, 2009.
Medicine, vol. 344, no. 12, pp. 873–880, 2001. [26] A. K. Heikhmakhtiar, A. J. Ryu, E. B. Shim et al., “Influence of
[11] J. G. Cleland, J.-C. Daubert, E. Erdmann et al., “The effect of LVAD function on mechanical unloading and electrome-
cardiac resynchronization on morbidity and mortality in chanical delay: a simulation study,” Medical and Biological
heart failure,” New England Journal of Medicine, vol. 352, Engineering and Computing, vol. 56, no. 5, pp. 911–921, 2017.
no. 15, pp. 1539–1549, 2005. [27] R. M. Delgado and B. Radovancevic, “Symptomatic relief: left
[12] A. J. Moss, W. J. Hall, D. S. Cannom et al., “Cardiac- ventricular assist devices versus resynchronization therapy,”
resynchronization therapy for the prevention of heart- Heart Failure Clinics, vol. 3, no. 3, pp. 259–265, 2007.
failure events,” New England Journal of Medicine, vol. 361, [28] J. Muratsu, M. Hara, I. Mizote et al., “The impact of cardiac
no. 14, pp. 1329–1338, 2009. resynchronization therapy in an end-stage heart failure pa-
[13] Y. Hu, V. Gurev, J. Constantino, and N. Trayanova, “Efficient tient with a left ventricular assist device as a bridge to re-
preloading of the ventricles by a properly timed atrial con- covery,” International Heart Journal, vol. 52, no. 4,
traction underlies stroke work improvement in the acute pp. 246-247, 2011.
response to cardiac resynchronization therapy,” Heart [29] H. Keilegavlen, J. E. Nordrehaug, S. Faerestrand et al., “Treat-
Rhythm, vol. 10, no. 12, pp. 1800–1806, 2013. ment of cardiogenic shock with left ventricular assist device
[14] R. John, F. Kamdar, K. Liao et al., “Improved survival and combined with cardiac resynchronization therapy: a case re-
decreasing incidence of adverse events with the HeartMate II port,” Journal of Cardiothoracic Surgery, vol. 5, no. 1, p. 1, 2010.
left ventricular assist device as bridge-to-transplant therapy,” [30] J. Constantino, Y. Hu, A. C. Lardo, and N. A. Trayanova,
Annals of Thoracic surgery, vol. 86, no. 4, pp. 1227–1235, 2008. “Mechanistic insight into prolonged electromechanical delay
[15] D. Casarotto, T. Bottio, A. Gambino, L. Testolin, and in dyssynchronous heart failure: a computational study,”
G. Gerosa, “The last to die is hope: prolonged mechanical American Journal of Physiology-Heart and Circulatory Phys-
circulatory support with a Novacor left ventricular assist iology, vol. 305, no. 8, pp. H1265–H1273, 2013.
device as a bridge to transplantation,” Journal of Thoracic and [31] Y. Hu, V. Gurev, J. Constantino, J. D. Bayer, and
Cardiovascular Surgery, vol. 125, no. 2, pp. 417-418, 2003. N. A. Trayanova, “Effects of mechano-electric feedback on
[16] M. A. Daneshmand, K. Rajagopal, B. Lima et al., “Left ven- scroll wave stability in human ventricular fibrillation,” PLoS
tricular assist device destination therapy versus extended One, vol. 8, no. 4, Article ID e60287, 2013.
criteria cardiac transplant,” Annals of Thoracic Surgery, [32] X. Jie, V. Gurev, and N. Trayanova, “Mechanisms of
vol. 89, no. 4, pp. 1205–1210, 2010. mechanically induced spontaneous arrhythmias in acute re-
[17] E. J. Birks, “Left ventricular assist devices,” Heart, vol. 96, gional ischemia,” Circulation Research, vol. 106, no. 1,
no. 1, pp. 63–71, 2010. pp. 185–192, 2010.
[18] E. A. Rose, A. C. Gelijns, A. J. Moskowitz et al., “Long-term [33] K. M. Lim and D. U. Jeong, “Influence of the KCNQ1 S140G
use of a left ventricular assist device for end-stage heart mutation on human ventricular arrhythmogenesis and
failure,” New England Journal of Medicine, vol. 345, no. 20, pumping performance: simulation study,” Frontiers in
pp. 1435–1443, 2001. Physiology, vol. 9, p. 926, 2018.
[19] G. T. Altemose, V. Gritsus, V. Jeevanandam, B. Goldman, and [34] V. Gurev, T. Lee, J. Constantino, H. Arevalo, and
K. B. Margulies, “Altered myocardial phenotype after me- N. A. Trayanova, “Models of cardiac electromechanics based
chanical support in human beings with advanced cardio- on individual hearts imaging data,” Biomechanics and
myopathy,” Journal of heart and lung transplantation: The Modeling in Mechanobiology, vol. 10, no. 3, pp. 295–306, 2011.
Official Publication of the International Society for Heart [35] J. Provost, V. Gurev, N. Trayanova, and E. E. Konofagou,
Transplantation, vol. 16, no. 7, pp. 765–773, 1997. “Mapping of cardiac electrical activation with electrome-
[20] H. R. Levin, M. C. Oz, J. M. Chen et al., “Reversal of chronic chanical wave imaging: an in silico–in vivo reciprocity study,”
ventricular dilation in patients with end-stage cardiomyop- Heart Rhythm, vol. 8, no. 5, pp. 752–759, 2011.
athy by prolonged mechanical unloading,” Circulation, [36] N. A. Trayanova, “Whole-heart modeling: applications to
vol. 91, no. 11, pp. 2717–2720, 1995. cardiac electrophysiology and electromechanics,” Circulation
[21] A. Barbone, M. C. Oz, D. Burkhoff, and J. W. Holmes, Research, vol. 108, no. 1, pp. 113–128, 2011.
“Normalized diastolic properties after left ventricular assist [37] N. A. Trayanova and J. J. Rice, “Cardiac electromechanical
result from reverse remodeling of chamber geometry,” Cir- models: from cell to organ,” Frontiers in Physiology, vol. 2,
culation, vol. 104, no. 1, pp. I-229–I-232, 2001. p. 43, 2011.
[22] S. Nakatani, P. M. McCarthy, K. Kottke-Marchant et al., “Left [38] K. M. Lim, J. Constantino, V. Gurev et al., “Comparison of the
ventricular echocardiographic and histologic changes: impact effects of continuous and pulsatile left ventricular-assist
devices on ventricular unloading using a cardiac electro-

mechanics model,” Journal of Physiological Sciences, vol. 62,
no. 1, pp. 11–19, 2012.
[39] K. ten Tusscher, D. Noble, P. Noble, and A. Panfilov, “A model
for human ventricular tissue,” American Journal of
Physiology-Heart and Circulatory Physiology, vol. 286, no. 4,
pp. H1573–H1589, 2004.
[40] O. Berenfeld and J. Jalife, “Purkinje-muscle reentry as
a mechanism of polymorphic ventricular arrhythmias in a 3-
dimensional model of the ventricles,” Circulation Research,
vol. 82, no. 10, pp. 1063–1077, 1998.
[41] J. J. Rice, F. Wang, D. M. Bers, and P. P. De Tombe, “Ap-
proximate model of cooperative activation and crossbridge
cycling in cardiac muscle using ordinary differential equa-
tions,” Biophysical Journal, vol. 95, no. 5, pp. 2368–2390, 2008.
[42] R. C. Kerckhoffs, M. L. Neal, Q. Gu et al., “Coupling of a 3D
finite element model of cardiac ventricular mechanics to
lumped systems models of the systemic and pulmonic cir-
culation,” Annals of Biomedical Engineering, vol. 35, no. 1,
pp. 1–18, 2007.
[43] J. B. Shea and M. O. Sweeney, “Cardiac resynchronization
therapy a patient’s guide,” Circulation, vol. 108, no. 9,
pp. e64–e66, 2003.
[44] Y. Hu, V. Gurev, J. Constantino, and N. Trayanova, “Opti-
mizing cardiac resynchronization therapy to minimize ATP
consumption heterogeneity throughout the left ventricle:
a simulation analysis using a canine heart failure model,”
Heart Rhythm, vol. 11, no. 6, pp. 1063–1069, 2014.

Mathematical Modeling and Models For Optimal Decision-Making in Health Care

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mathematical Modeling and Models For Optimal Decision-Making in Health Care

Uploaded by

Copyright:

Available Formats

Computational and Mathematical Methods in Medicine

Mathematical Modeling and Models

Mathematical Modeling and Models for Optimal

Lead Guest Editor: Giedrius Vanagas

An Improved Sliding Window Area Method for T Wave Detection

A Technical Review of Convolutional Neural Network-Based Mammographic Breast Cancer Diagnosis

Automatic Segmentation of Pathological Glomerular Basement Membrane in Transmission Electron

Minimalistic Approach to Coreference Resolution in Lithuanian Medical Records

Evaluation of the Feasibility of Screening Tau Radiotracers Using an Amyloid Biomathematical

Continuous Blood Pressure Estimation Based on Two-Domain Fusion Model

Giedrius Vanagas ,1 Tomas Krilavičius ,2,3 and Ka Lok Man 4,5

Correspondence should be addressed to Giedrius Vanagas; g.vanagas@gmail.com

Received 31 July 2019; Accepted 1 August 2019; Published 14 August 2019

[8] Y.-H. Nai, M. Shidahara, C. Seki, and H. Watabe, “Bio-

Bernhard Haller , Kurt Ulm, and Alexander Hapfelmeier

Correspondence should be addressed to Bernhard Haller; bernhard.haller@tum.de

Received 1 February 2019; Accepted 22 May 2019; Published 13 June 2019

Guest Editor: Tomas Krilavičius

1. Introduction growth factor receptor (EGFR) mutation in non-small cell

2.2. Simulation Settings. Data were generated to mimic data

Consequently, the hazard ratio is close to one for small

1.75z8 Interaction (MFPI) procedure with the FP1-ﬂex3 strategy

Low censoring High censoring

Median split Median split

Median split Median split

Median split Median split

Sample size Sample size

Median split Median split

Median split Median split

Median split Median split

Sample size Sample size

Low censoring High censoring

Median split Median split

Median split Median split

Median split Median split

Median split Median split

Median split Median split

Median split Median split

Low censoring High censoring

Median split Median split

Median split Median split

Median split Median split

Median split Median split

Median split Median split

Median split Median split

Data Availability [13] R. Wang, S. W. Lagakos, J. H. Ware, D. J. Hunter, and

Li Luo ,1 Jialing Li ,1 Xueru Xu ,1 Wenwu Shen ,2 and Lin Xiao 1

Guest Editor: Giedrius Vanagas

1. Introduction Amongst all of the medical institutions, China’s large ter-

Data LOS Department

Bed Arrival rate

Quadratic curve function

Figure 2: Methodology: process of the data-driven hybrid three-stage framework.

Kij ≥ 2, i � 1, 2, . . . , n, j � 1, 2, 3. (8) 3.2. Stage II: Construction of Constraint Conditions. After

Table 2: Department information summary.

Table 3: Results of arrival rate distribution and LOS.

Figure 3: Simulation model of the W9.

Table 9: Relationship between bed numbers and BOR in W9.

Table 10: Relationship between bed numbers and BOR in W10.

Table 11: Relationship between beds and BOR in W6.

Table 12: Relationship between beds and BOR in W27.

Table 13: Relationship between beds and BOR in W19.

Received 4 December 2018; Accepted 5 March 2019; Published 1 April 2019